The Comparative Toxicogenomics Database facilitates identification and understanding of chemical-gene-disease associations: arsenic as a case study

Curation process

Curated interaction analysis

All curated interaction data in CTD are publicly available. The arsenic analysis reported here is based on data downloaded and analyzed in March 2008. Prior to analysis, the data set was first manually processed as follows:

1. Curated interactions for arsenic compounds were combined. Interactions for both arsenic and arsenicals were retrieved using the CTD Curated Interaction Query [15]. The query was a "contains" search using the term "arsenic" in the Chemical query box and "ANY" as the chemical-gene interaction type. This query retrieved all curated interactions for arsenic, arsenicals, and all child terms.

2. "No effect" interactions were removed. In addition to capturing chemical-gene interactions, curators capture "no effect" results reported in the literature, such as "arsenic trioxide does not affect the phosphorylation of protein A." To avoid artificially inflating the number of interactions involving arsenic or falsely associating arsenic with a gene, all "no effect" interactions were removed from the data set.

3. Binary interactions were identified. A curated interaction in CTD must include at least one chemical and one gene (binary interaction); however, complex interactions describing more than two molecules may also be captured. For example, the following two interactions were curated from the same paper: "arsenite results in increased expression of TP53 mRNA" and "alpha-tocopherol inhibits the reaction [arsenite results in increased expression of TP53 mRNA]." The first interaction is binary and describes the core data (arsenite exposure leads to up-regulated TP53 mRNA). The second interaction is complex in that it reiterates the core data and then qualifies it with additional information. To avoid inclusion of duplicate interactions from a single reference, only binary interactions were retained in the data set.

CTD controlled vocabularies

Several controlled vocabularies and ontologies are used to construct chemical-gene interactions and chemical- and gene-disease relationships, allowing data to be captured consistently by different curators, retrieved more easily and reproducibly by users, and integrated with other databases. Biocurators use official terms (gene symbol, chemical name, or disease name) from standardized sources (below). CTD is updated monthly, using the most current release of all external databases that are integrated with CTD (see "Data Status" link on CTD homepage for specific release versions). If an official term undergoes a nomenclature change, CTD implements that same change. Consequently, CTD is synchronized regularly with its external data sources.

1. Chemicals. The CTD chemical vocabulary derives from a modified subset of the chemicals and supplementary concepts in the "Drugs and Chemicals" category of MeSH.

2. Chemical qualifiers. A chemical term can be qualified with "analog" or "metabolite" to describe a chemical derivative (e.g., arsenic disulfide analog or arsenic metabolite).

3. Diseases. The CTD disease vocabulary uses terms from both MeSH and OMIM. To utilize both sources, we first integrated OMIM disease names (a flat list) into the MeSH disease hierarchy. Most OMIM diseases easily mapped to a single equivalent term from MeSH (e.g., OMIM "lung cancer" maps directly to MeSH "lung neoplasm"). Some OMIM disease names were mapped to multiple MeSH terms when a single equivalent term was not available (e.g., OMIM "chronic myeloproliferative disorder with eosinophilia" was most accurately mapped to two MeSH terms: "myeloproliferative disorders" and "eosinophilia"). This mapping preserved individual diseases but rooted them in a hierarchical vocabulary, allowing a user to cluster what might at first appear to be unrelated diseases (e.g., "Papillon-Lefevre syndrome" and "multiple café-au-lait spots") via a more general parent disease (e.g., "skin diseases").

4. Genes. CTD uses official gene symbols and names from the National Center for Biotechnology Information's (NCBI) Entrez-Gene database [16].

5. Gene qualifiers. A gene symbol can be qualified with any of 15 gene qualifier terms developed at CTD (e.g., "DNA," "promoter," "mRNA," "protein," etc.) to specify any aspect of a gene. Henceforth, we use the word "gene" in this paper to collectively refer to any aspect of a gene, including its product.

6. Action codes. CTD curators developed a hierarchical vocabulary of 50 diverse terms (e.g., "binding," "phosphorylation," "activity," "transport," "methylation") to describe the specific molecular interaction between any chemical and gene. A complete list of action codes and their definitions is available via the Help link for interactions on CTD glossary page and gene, interaction, and reference query forms.

7. Organisms. The CTD organism vocabulary consists of the Eumetazoa portion (vertebrates and invertebrates) of the NCBI Taxonomy database [17].

Gene function and network analysis

GO and KEGG pathway data are integrated with CTD gene data. Annotations were downloaded for arsenic-interacting genes using the CTD batch query tool. Network analyses were performed using the Ingenuity Pathways Analysis (Ingenuity Systems) software.

Disease analysis

There are two sources for direct disease relationships in CTD: curation of chemical-disease and gene-disease relationships from the literature, and integration of gene-disease relationships from the OMIM database [9]. Disease relationships in CTD may be direct or inferred. Direct relationships are established by CTD curation and imported from OMIM. Inferred relationships are generated computationally by integrating curated chemical-gene data with curated gene-disease data. For example, if Chemical A interacts with Gene B (via a curated chemical-gene interaction), and independently Gene B is associated with Disease C (via a curated gene-disease molecular mechanism or biomarker relationship), then Chemical A is said to have an inferred relationship with Disease C (inferred via Gene B). Disease relationships are clearly labeled as either direct or inferred in CTD. Of the 1,456 genes that interact with arsenic, 424 genes have a direct mechanism or biomarker relationship with 516 unique diseases. To identify whether there was a pattern among these 516 diseases, the diseases were rooted to their top-leveled MeSH terms and clustered into 26 categories. For example, type 1 diabetes mellitus is associated with three paths of the MeSH disease hierarchy that ultimately are rooted to three top-level disease categories: metabolic diseases, endocrine system diseases, and immune system diseases. Consequently, any arsenic-interacting gene directly associated with type 1 diabetes mellitus is counted in all three of those disease categories. The final distribution pattern for these disease categories was compared against the disease category distribution pattern for all disease-associated genes in CTD. A Pearson's chi-square test comparing arsenic associated diseases (derived from 424 arsenic-interacting genes) indicated that these associations were highly significant (p-value < 0.0001).

Arsenic compounds

Sodium arsenite, arsenic trioxide, and sodium arsenate are the three most frequently curated arsenic compounds in CTD, interacting with 793, 500, and 230 genes, respectively (Table 1). Despite the complex chemistry and biotransformation of arsenic compounds, many (e.g., sodium arsenite and arsenic trioxide) are metabolized into a common reactive arsenite ion [18]. To determine whether different arsenic compounds affect similar genes, we compared the interacting gene sets for these top three arsenic compounds. Collectively, sodium arsenite, arsenic trioxide, and sodium arsenate had curated interactions with 1,297 genes, accounting for 89% of the total genes analyzed in this arsenic study. Among these genes, 206 overlapped with at least two compounds, with the least overlap found between sodium arsenate and arsenic trioxide (Figure 1). Twenty genes interacted with all three arsenic compounds. These data indicate that arsenic compounds share some common molecular activity, although the overall biological effects or the molecular networks they invoke may be compound-specific. Analysis of all 1,456 genes associated with all 21 arsenic compounds accounts for primary and secondary targets of arsenic and its metabolites. This approach may reasonably mimic a biological response to arsenic. It is the combination of primary, secondary and metabolite effects that contribute to the overall response to an exposure.

Molecular actions of arsenic

Curated chemical-gene interactions in CTD indicated that arsenic has a broad range of molecular effects (Table 2). The most common arsenic interaction involved effects on transcription: 810 genes were up-regulated and 652 were down-regulated. Arsenic also affected protein activity, phosphorylation status, degradation, localization, secretion, metabolism, and stability. Promoter methylation and gene mutagenesis effects of arsenic were also described. Molecular interactions involving effects of proteins on arsenic included resistance/susceptibility to arsenic, arsenic metabolism, arsenic transport, and arsenic binding.

Arsenic-interacting genes

Gene Ontology (GO) annotations and KEGG pathway data are integrated in CTD to provide additional biological context to genes and chemical effects. GO annotates gene products in terms of their associated molecular function, biological process, and cellular component [7]. KEGG is a public resource that presents information about molecular reaction networks [8]. Arsenic-interacting genes were evaluated using GO, KEGG, and Ingenuity Pathway Analysis (IPA) to gain insight into the biological functions and molecular networks that may be altered by arsenic exposure.

The top 25 genes with curated arsenic interactions included several transcription factors (JUN, MYC, TP53, EGR1, ESR1, DDIT3, FOS) and regulators of cell cycle, cell proliferation, and apoptosis (BCL2, CCND1, CDKN2A, VEGFA, CDNK1A, GADD45A, BAX), suggesting that a dramatic cellular response may be induced by arsenic exposure (Table 3). Seven of these genes (HMOX1, CCND1, CDKN2A, GADD45A, GSTP1, DDIT3, FOS) interact with multiple arsenic compounds (Figure 1).

GO and KEGG pathway annotations for all arsenic-interacting genes in CTD were evaluated and comprised 1,042 distinct molecular functions, 2,093 biological processes, 383 cellular components, and 182 pathways (Table 4). The most prominent annotated terms from each category included protein binding (GO: molecular function); transcription or signal transduction (GO: biological processes); nucleus, cytoplasm, or membrane (GO: cellular component); and MAPK signaling pathway (KEGG). Other frequently annotated processes and pathways for the 1,456 genes included apoptosis, cell cycle, cell proliferation, immune response, and cell signaling (Table 4). These processes and pathways were also implicated when using only the top 25 genes with curated arsenic interactions (Table 3), suggesting that this subset of genes may accurately represent the overall biology of genes responding to arsenic exposure.

To provide a more detailed view of molecular networks potentially affected by arsenic, IPA was used to analyze arsenic-interacting genes in CTD. Analysis of the 1,376 genes with annotations in the IPA database identified a large arsenic-modulated interaction network comprising 105 proteins (Figure 2; see additional file 1). Pathway analysis showed that the network was enriched with roles in cell cycle control, apoptosis, and DNA repair, and was associated significantly with cancer. The network contained transcriptional regulators (28 proteins), enzymes (18 proteins), cytokines (4 proteins), ligand-dependent nuclear receptors (4 proteins), transmembrane receptors (3 proteins) and peptidases (2 proteins). Locations of the proteins within the network were distributed among the nucleus (52 proteins), cytoplasm (27 proteins), extracellular space (8 proteins) and plasma membrane (8 proteins). Biological implications of this network are consistent with GO, KEGG, and disease associations identified by other methods and described in this report.

Arsenic-gene-disease associations

Curated direct chemical-gene interactions and direct gene-disease associations in CTD are integrated to create inferred chemical-disease relationships. For example, if Chemical A interacts with Gene B (via a curated chemical-gene interaction), and Gene B is associated independently with Disease C (via a curated gene-disease relationship), then CTD presents an inferred relationship between Chemical A and Disease C (inferred via Gene B). We assessed whether it was valid to use these inferred relationships to predict diseases influenced by environmental chemicals.

Among the genes with curated arsenic interactions in CTD, 424 have curated relationships with 516 diseases. Therefore, arsenic has inferred relationships with 516 diseases in CTD. Using the hierarchical disease vocabulary in CTD, these diseases were clustered into 26 general categories (see Methods). The most common disease categories associated with arsenic included neoplasms, nervous system diseases, skin diseases, digestive system diseases, metabolic disorders, and immune system diseases (Table 5). Notably, many of these diseases were corroborated in the literature as being associated with arsenic, including malignancies (skin, lung, liver, kidney and bladder cancer), neurological defects (peripheral neuropathy and cognitive impairment), skin lesions, diabetes, and many others (Figure 3) [19-21]. This corroboration supported the value of using CTD-inferred disease relationships to develop novel hypotheses about the health effects of exposure to environmental chemicals.

CTD provides candidate genes that may help to explain mechanisms underlying chemical-disease relationships. Cancer was the disease category most frequently associated with arsenic via curated interactions with 177 genes in CTD. These diseases were subdivided into specific types of malignancies, each with a unique set of associated genes: breast (49 genes), stomach (42 genes), leukemia (23 genes), nervous system (19 genes), lung (18 genes), colon (17 genes), bladder (15 genes), and others (Table 6). Although some of these disease associations have been documented (e.g., lung cancer), the molecular mechanisms underlying the etiologies remain largely unclear [22-24]. The inferred relationship between arsenic and lung cancer is established in CTD via curated interactions with 18 genes. Notably, IPA analysis of these genes identified a network involving 72% of these genes (AREG, BIRC5, EGFR, ERBB2, ESR1, GCLC, HMOX1, KRAS, MAP3K8, MAPK14, PIK3CA, RASSF1, TGFA). This network suggests that these genes may function coordinately and form a basis for investigating a mechanistic connection between arsenic and lung cancer. CTD enables analyses that will lead to the development of novel hypotheses about the molecular etiologies of environmental diseases.