In this analysis, we used available data in NCBI databases, plus three software suites (DAVID, PDG-ACE, and GeneGo). To identify potential candidate genes, we queried the Entrez Gene database 11 for "depression AND ethanol AND human [orgn]". Since Entrez Gene surveys the whole human genome, this approach is complementary to Whole Genome Association (WGA) analysis 121314. WGA surveys many genetic variants (e.g. ~1 million Single Nucleotide Polymorphisms (SNPs)) across the genome, while this text-based query surveys all genes in Entrez Gene. WGA provides indirect information on potential associations of genes with disease by associating nearby SNP variants with disease, while Entrez Gene queries associate genes with disease based on curated text describing the genes. Each approach has advantages and disadvantages, highlighted in Table 1, though we believe that the query-based approach offers significant advantages in the initial phase of candidate gene analysis. First, in recent years Entrez Gene has become a rich source of information on many genes, while WGA typically focuses on a very limited information source (e.g. SNP association). Second, data provided by NCBI are free to the users. While some of the genotypic and phenotypic data required for WGA are becoming available in free public repositories, these data are generally expensive to acquire. Third, data derived from curated sources such as Entrez Gene are backed by multiple rounds of experimentation and peer review, so the believability of a positive result, as measured by Positive Predictive Value [PPV = True Positives/(True positives + False Positives)], is very high. Given these strengths, a query based approach should be considered for the initial phase of candidate gene analysis, which will then maximize the value of subsequent association testing.

The Entrez Gene database provides reliable information about most genes, but the Entrez Gene record is a summary, so the information it provides is necessarily limited. Therefore, we queried a second NCBI database, PubMed 15, for more complete information on a variety of hypotheses developed in this analysis. Table 2 shows the short form of each PubMed query in column 1 and the citation count is in column 2. For each PubMed query, the actual query was qualified by Medical Subject Headings (MeSH) annotation 16 and the detailed queries are included in Additional file 1. Publications are annotated by MeSH terminology to provide a consistent way to retrieve information in publications that may use different terminology for the same concepts. In addition, querying PubMed by MeSH annotation yields a high likelihood that publications returned by the query refer to the queried terms in the appropriate context (high PPV). For example, in looking for association between depression and AUD, we queried PubMed for "("Depressive Disorder" [Mesh] OR "Depression" [Mesh] OR "Depressive Disorder, Major" [Mesh]) AND ("Ethanol" [Mesh] OR "Alcohol-Induced Disorders, Nervous System" [Mesh])". The resulting 263 citations returned are consistent with the statistically significant association between depression and AUD found by Hasin and Glick 2. Using the University of Michigan's electronic journals library, we made an exhaustive review the 263 citations returned by this query to quantify the PPV of this MeSH qualified PubMed query. In this review, we looked for clear evidence that the authors were studying AUD and depression, in English language publications, either full text or abstracts. We were able to unequivocally assess the accuracy of MeSH annotation for AUD and depression in 179 citations. Of these 179 publications, 171 clearly refer to both AUD and depression, though not necessarily the comorbidity of AUD with depression. Of the eight remaining publications, five are correctly annotated for ethanol, but ethanol is used as an important reagent in the experimental procedures. The final three publications are incorrectly annotated for ethanol, though they refer to pharmacologically related chemicals (flupenthixol, haloperidol, and pirbuterol). We consider these eight publications to be false positives with respect to our original query, yielding a 95.5% (171/179) PPV for our query.

There are, however, limitations to the use of MeSH annotation in querying publications. First, while MeSH contains thousands of Medical Subject Headings, it is not an exhaustive annotation of all potential biomedical contexts. Second, there may be missing annotation or ambiguity in the annotation for some publications. For example, as mentioned in the PPV analysis, not all of the 171 papers that we counted as True Positives were specifically about comorbid AUD with depression. The earlier papers tended to refer to AUD and depression as co-occurring phenotypes, while the later papers focused on evaluating the comorbidity. There also may be ambiguity in the user's choice of MeSH terms queried, and some publications may report negative associations, though the bias against publication of negative results minimizes the number of these papers. And finally, a lack of citations is not evidence of a lack of association. While co-occurrence of queried terms is not proof of association, co-occurrence does provide a quantitative measure of peer reviewed research on the association of these specific terms and complements other forms of data that can be mined. In future analyses, Natural Language Processing will help to improve these queries by mining the full text of publications for a wide variety of concepts, and provide additional evidence of association among terms. Given believable candidate genes, derived from queries of available data at Entrez Gene and PubMed, we continued the analysis using DAVID, PDG-ACE, and MetaCore (GeneGo, Inc.) to understand how these candidate genes may be related to the etiology of AUD, depression, and comorbid AUD with depression.

DAVID software, the PDG-ACE algorithm, and the GeneGo suite perform related tasks. The essential assumption in the use of all three tools is that, in complex diseases, multiple genetic influences converge on a single phenotype. The three tools use different approaches to identify how multiple genes may influence the phenotype of interest and assess the statistical significance of the results seen.

DAVID 17 uses a database of available gene annotation to overlay candidate genes onto predefined gene sets. For example, genes may be organized into predefined sets based on annotation for their participation in a Gene Ontology function, process, or component 18. A set of candidate genes is then derived from a genotype/phenotype analysis (e.g. WGA, microarray expression, or text-based database query) and the candidate genes are overlaid on the predefined gene sets. If a large proportion of the candidate genes are found to be annotated for a given gene set (e.g. an over-represented Gene Ontology process), the candidate genes may have an important impact on the process represented by that gene set. Since the candidate genes were initially derived from a genotype/phenotype relationship, the over-represented process may be important in the phenotype 19.

However, since much gene annotation remains incomplete, PDG-ACE 20 [see Additional file 2] serves as an adjunct to software that depends on gene annotation. As noted, the Entrez Gene database offers a reliable summary of information on a wide range of genetic and environmental influences for each gene (e.g. gene function, disease influences, localization, environmental effects). However, much of the information available in Entrez Gene is free text, rather than formal annotation, limiting the use of tools like DAVID. We created the PDG-ACE algorithm to overcome this limitation by mining Entrez Gene text to identify biomedical keywords that are common and significantly over-represented in the descriptions of genes at a selected locus pair, relative to all locus pairs in the genome.

The PDG-ACE algorithm uses a controlled vocabulary of biomedical keywords to limit the search to concepts likely to be useful in understanding disease etiology and show statistically significant over-representation. For a given locus pair, PDG-ACE mines the Entrez Gene text to find keywords from the controlled vocabulary that are common across the locus pair. For keywords that are common across the locus pair, significance of over-representation is established by permutation testing. Keywords that are common and significantly over-represented across the locus pair highlight rare combinations of genes that share the biomedical concepts associated with the keywords. Testing of the PDG-ACE algorithm, using both positive controls (documented gene-gene interactions in complex disease) and negative controls (randomly selected gene pairs), demonstrates that PDG-ACE provides insight into the etiology of true genetic interactions, while excluding loci where the data is insufficient to identify interactions 20. As with DAVID, the loci are initially defined by the disease phenotype, so the over-represented keywords may offer insight into the nature of a genetic interaction in disease etiology. In applying the PDG-ACE algorithm, we define genetic interaction to mean a "statistically significant multi-gene or multi-gene-by-environment influence on the phenotype", consistent with Hartman et al. 21. This definition is deliberately broad, allowing PDG-ACE to identify unannotated influences on the phenotype, ranging from well-defined influences (e.g. epistasis, protein-protein binding, canalization, genetic robustness, buffering) to completely novel influences, as long as the influence is multi-gene (or multi-gene-by-environment) and statistically significant. The complete list of keywords used in this analysis is included in Additional file 3. Software implementing the PDG-ACE algorithm is available on request from Dr. Ben Keller, bkeller@emich.edu.

The GeneGo suite 22 uses the MetaCore database of documented gene-gene (protein-protein) interactions 23, as well as a set of graphical and statistical tools, to allow researchers to build gene networks based on relationships among genes. The MetaCore interactive database has been manually curated from publications describing interactions among proteins and small molecules of biological relevance in humans. We used GeneGo to place our candidate genes into a cellular context, evaluate the significance of gene networks that our candidates participate in, and build a model that integrates all of the information derived in our analysis. Specifics of the parameter settings used in the GeneGo analysis are included in Table 3. Notably, we accepted only direct interactions, curated by human reviewers.

We first searched for evidence of genetic influences on depression and AUD by querying NCBI's Entrez Gene database for "depression AND ethanol AND human [orgn]". This query returned exactly three genes: Tumor Necrosis Factor (TNF superfamily, member 2, GeneID 7124), Methylenetetrahydrofolate Reductase (MTHFR, 5,10-methylenetetrahydrofolate reductase (NADPH), GeneID 4524) 242526272829, and APOE (apolipoprotein E, GeneID 348). However, the Gene Ontology annotation for APOE, "response to ethanol", is inferred from electronic annotation, rather than manual curation. In an effort to use only the most reliable information to establish our initial candidate genes, we excluded APOE from the formal analysis. However, we have included a brief analysis of APOE in the results.

We searched PubMed for supporting information on potential roles of the proteins coded by TNF and MTHFR in depression, AUD, and comorbid depression with AUD. In searching PubMed, we found evidence that both TNF and MTHFR have been associated with depression and with ethanol in the literature, but not with the comorbidity of depression with AUD (Table 2). Notably, the result in Table 2 showing that TNF is associated with both depression and ethanol in a single publication is one of the false positive results identified in the PPV analysis. We didn't find evidence of MTHFR being associated with both depression and AUD in any single paper.

In total, the published literature is consistent with the hypothesis that TNF and MTHFR may both influence, or be influenced by, depression and AUD. We consider TNF and MTHFR to be valid candidate genes for both depression and AUD. Given these candidate genes, we next sought evidence that they participate in the comorbidity via genetic interaction.

Given evidence of genetic influences on depression and AUD, we refined the hypothesis to include interaction between TNF and MTHFR. We first used DAVID software to explore functional annotation consistent with interaction between TNF and MTHFR. DAVID looks for over-representation of candidate genes in signaling or metabolic pathways, Gene Ontology processes, etc., based on available gene annotation. TNF and MTHFR were not found to be significantly over-represented in any gene set tested by DAVID.

Since most genes are only partially annotated, we next explored potential interactions between TNF and MTHFR via the PDG-ACE algorithm 20. A total of 149 keywords were found to be common across the TNF/MTHFR gene pair, out of 2,531 keywords in this controlled vocabulary. We checked these 149 common keywords for over-representation using 10 million iterations in permutation testing. After a Bonferroni correction for 2,531 hypothesis tests, 27 keywords were found to be significantly over-represented (p-value < 0.05) in the TNF/MTHFR gene pair, relative to randomly selected gene pairs in the human genome. Six of these 27 keywords were found to be used in different contexts in the TNF and MTHFR Entrez Gene records and were eliminated; leaving 21 keywords that are common and significantly over-represented across the TNF/MTHFR gene pair (Table 4). Given this evidence of interaction between TNF and MTHFR in AUD and depression, and consistent with expected environmental influences on complex diseases, we further refined the hypothesis to include gene-by-environment interaction influencing the comorbidity.

Each significantly over-represented keyword identified by PDG-ACE represents one hypothesis on the etiology of a genetic interaction between TNF and MTHFR in depression and AUD. We tested each of these hypotheses by querying PubMed for "depression AND ethanol AND keyword" (Table 4). As in the previous round of testing, we used MeSH annotation to limit the queries and we used citation counts to provide indications of relevant association in the literature. Notably, since ethanol is a keyword in the PDG-ACE controlled vocabulary that we used, it acts as a positive control in this analysis and is the top keyword when ranked by relevance. In addition to ethanol, the top keywords ranked by relevance to depression and AUD are "consumption", "background", and "intake". In the Entrez Gene records for TNF and MTHFR, intake and consumption both refer to intake of alcohol, while background refers specifically to genetic background. Results of the PDG-ACE analysis are consistent with the hypothesis that genetic background, via TNF and MTHFR, as well as environmental influences, via alcohol intake or consumption, are interacting elements of susceptibility in comorbid depression with AUD. Given evidence of genetic interaction between TNF and MTHFR, as well as gene-by-environment interaction in comorbid depression with AUD, we next sought to understand TNF and MTHFR in a cellular context.

We used the GeneGo suite 2223 to place TNF and MTHFR in a cellular context. Results of the PDG-ACE analysis suggested that ethanol consumption exerts an environmental influence on this genetic system. Also, TNF is a secreted protein that responds to the environment by binding TNF receptors to influence intracellular signal transduction pathways that regulate gene expression. Given these inputs, we focused our GeneGo analysis on cell signaling via signal transduction and the regulation of gene expression within the cell (input parameters in Table 3).

The most significant network found by GeneGo (Figure 2) shows a feedback loop among TNF (a.k.a. TNF-alpha), TNF-R1 (a.k.a. TNFRSF1A, GeneID 7132), and NF-kB (a.k.a. NFKB1, GeneID 4790). Note that these proteins do not act in isolation, but the GeneGo graphic shows the essential elements of the network based on annotation in GeneGo's MetaCore database. In the network identified by GeneGo, conditions in the extracellular environment are sensed inside the cell via binding of TNF to TNF-R1, activating TNF-R1 3031. Activated TNF-R1 activates NF-kB 3233, which subsequently activates the expression of both TNF 34 and MTHFR 3536. Given the potential for TNF-R1 and NF-kB to impact this network, we queried PubMed for potential influences of these genes on depression and AUD (Table 2). We found citations consistent with roles for both of these genes in depression and AUD.

Genetic variation or extracellular signals that affect any part of the feedback cycle (Figure 2) would tend to be amplified over a number of cycles, potentially leading to a growing imbalance in the system over time. In this network, NF-kB activates the expression of MTHFR, which metabolizes folate, so imbalances in the system would be expected to lead to imbalances in folate metabolism. This result is consistent with evidence of decreased folate levels in alcoholism 37. In addition, reduced folate levels are associated with depression 38, where research on etiology suggests that reduced folate levels may alter neurotransmitter metabolism 394041.

Summarizing our results to this point, we found evidence that both TNF and MTHFR are associated with both depression and alcohol in the literature, TNF and MTHFR interact to influence the comorbidity and alcohol exerts an environmental influence on this genetic network, and folate levels are altered in both depression and alcohol dependence. Investigating the environmental influence of ethanol on TNF, we found that ethanol intake suppresses the function of TNF Alpha Converting Enzyme (TACE, a.k.a. ADAM17, GeneID 6868) 28, leading to reduced TNF signaling. Integrating these lines of evidence, we propose a model of comorbid depression with AUD (Figure 3). In this model, environmental ethanol exposure suppresses TNF signaling by inhibiting TACE 28. Reduced TNF signaling reduces TNF-R1 activation 3031, which then reduces NF-kB activation 32. NF-kB activates the expression of both TNF 3334 and MTHFR 3536, so reduced NF-kB activation contributes to further reduction in TNF signaling as well as reduced MTHFR expression. Reduced expression of MTHFR would be expected to alter folate metabolism and increase susceptibility to depression 38.

Signal transduction, leading to transcriptional regulation of both TNF and MTHFR, is an essential assumption of this model. In addition to building networks, GeneGo analyzes networks for over-representation of genes that participate in Gene Ontology (GO) processes (similar to the DAVID analysis described), then ranks the GO processes by p-value. Table 5 shows that "positive regulation of transcription from RNA polymerase II promoter" is the most significantly over-represented GO process in this network. This result is consistent with the hypothesis that environmental influences, via ethanol exposure, are transduced through this genetic network to alter TNF and MTHFR expression, folate metabolism, and susceptibility to depression.

Apolipoprotein E (APOE) was excluded from the formal analysis because the annotation in Entrez Gene was inferred, rather than curated. In pursuing the hypothesis that APOE is involved in comorbid depression with AUD, we re-started the analysis including APOE. We queried PubMed for citations consistent with this hypothesis (Table 2) and found evidence of APOE's role in both depression and AUD, but no evidence for a role in the comorbidity. PDG-ACE analysis did not produce any significant results after correction for multiple hypothesis tests. We attempted to insert APOE into the GeneGo network (Figure 2) and found no annotated interactions between APOE and the other genes in the network. Apolipoprotein E may well have an influence on comorbid depression with AUD but, if so, the mechanism is either independent of the network modeled or, more likely, the data is not yet available to reliably connect APOE to this network.

Implications of this model are consistent with dietary guidance recommending monitoring and appropriate supplementation of folate for patients in treatment for depression 484950, AUD, or comorbid depression with AUD 3738424344. In addition, identification of the variants associated with risk could be useful in prognosis and treatment of either or both conditions 51. Pharmacogenomic approaches are being successfully implemented in psychiatry, to the great benefit of patients with specific genetic variants 525354, and the identification of disease predisposing variants will likely improve the prognosis for depression with AUD.