DiseaseConnect: a comprehensive web server for mechanism-based disease–disease connections

Disease annotations

To construct systematically disease–gene and disease–drug relations using multiple databases, we used the unified medical language system (UMLS) that provides a comprehensive set of medical concepts and DO terms and includes a metathesaurus (24). We selected two vocabularies: Medical Subject Headings (MeSH) and the UMLS Metathesaurus. To describe disease–gene relations, DiseaseConnect uses only UMLS concepts with one of the following disease-related semantic types: ‘Pathologic Function’, ‘Injury or Poisoning’ and ‘Anatomical Abnormality’. To support a wide range of diseases, we combined all of the disease concepts from gene expression data, OMIM and GWAS databases. To avoid too general concepts, we only used those UMLS concepts that are associated with < 50 descendant concepts. These criteria result in a universe of 4791 disease concepts.

Disease-related gene expression

We collected a total of 1366 human gene expression datasets on 27 February 2014 from the Gene Expression Omnibus (GEO) (25), which consist of all 1355 human GEO datasets (GDS) that are pre-processed by GEO, and additionally 11 GEO series (GSE) processed by us as part of an NIH-funded project studying heart, lung, blood and sleep disorders. If one gene has several probes, we used the mean expression value for the gene. Within each sample, the expression values of all genes were log-transformed, median-centered and normalized to have unit standard deviation. The 1366 datasets contain in total 6024 subsets and 30 300 samples. Each subset is a group of samples with specific phenotypic traits or treatments. These subsets can be classified into different types, e.g. disease state, agent, cell line, cell type, tissue, protocol and infection. To concentrate on the disease studies, we selected 1178 subsets with the ‘disease state’ type. We manually assigned UMLS concepts to subsets, and then retrieved their parental UMLS concepts to complete the UMLS annotation for all subsets. To identify differentially expressed genes (DEGs) in each dataset, we selected all disease subsets with at least three samples and manually identified the normal subset in the dataset. We then performed a t-test between two subsets without overlapping samples. DEGs are defined as the top 100 genes with t-test P-values < 1e−6 for each subset pair. This analysis identified 928 subset pairs with DEGs. Overall, the 1366 gene expression datasets contributed 173 992 disease–gene relations, including 505 diseases and 11 812 genes.

Disease-related microRNA expression

We collected 233 human microRNA expression datasets from the GEO, which consist of 5194 samples. Each dataset has multiple phenotype-specific subsets. For studying disease-related microRNA data, we selected 84 datasets with disease and normal subsets. To systematically process phenotype information for microRNA expression data, we transferred the text description of microRNA datasets to UMLS concepts using the MetaMap program (26). To identify differentially expressed microRNAs in each dataset, we selected the disease subsets with at least three samples as well as the normal subset. We then performed a t-test between the normal subset and each disease subset without overlapping samples. We obtained the differentially expressed microRNAs with t-test P-values < 0.001 for each subset pair. This analysis resulted in 956 differentially expressed microRNAs identified in 73 subset pairs. The microRNA expression data contributed 17 088 disease–microRNA relations, including 20 diseases and 745 microRNAs.

Disease-related SNP

We downloaded the catalog of published GWAS from the National Human Genome Research Institute (27). This data is a manual curation of published GWAS hits, which may contain causative SNPs (27). The GWAS catalog contains descriptions of diseases/traits, associated SNPs, and reported genes. We converted these disease annotations to UMLS concepts using the MetaMap program (26) and selected GWAS genes with P-values < 1e−6. The GWAS data contributed 21 865 disease–gene relations, including 719 diseases and 3397 genes.

Disease–drug relationships

We extracted the disease–drug annotations and relationships from the UMLS National Drug File Reference Terminology (NDFRT). This database contributed 113 498 drug–disease relations, including 8995 drugs and 1525 diseases.

Disease comorbidity relationships

We downloaded disease comorbidity relations from the Phenotypic Disease Network (PDN) (1), which uses ICD9 disease codes. We selected disease–disease connections with t-value > 1.96 (P < 0.05), and then translated the ICD9 disease codes into UMLS concepts using the UMLS Metathesaurus. The PDN data contributed 132 786 disease–disease connections, including 2991 diseases.

Drug–gene relations

We downloaded 12 617 human drug–gene relations from the DrugBank database (28). These data included 3417 drugs and 1481 genes.

Disease–gene relationships from GeneRIF

GeneRIF is a database archiving short statements about the functions and disease relevance of genes. Each GeneRIF statement has a pointer to the PubMed ID of a scientific publication that provides evidence for the statement. We downloaded the GeneRIF database from the NCBI Gene database, and then for each gene we extracted UMLS disease concepts from its GeneRIF statements using the UMLS natural language processing tool, MetaMap program (26). This source provided 276 089 disease–gene relations, including 3113 diseases and 11 190 genes.

Disease–gene relationships from literature corpus mining (GeneWays)

We extracted all human disease–gene relationships from the GeneWays system (version 8.0) that we developed previously (29). This system has automatically extracted, analyzed, visualized and integrated molecular pathway data in 483 107 full-text articles and 11 826 299 abstracts downloaded from PubMed. We obtained 170 329 disease–gene relations, including 2046 diseases and 7239 genes. This set of disease–gene relationships has less than 20% overlap (33 137 shared) with those derived from GeneRIF. This difference in the two derived knowledge sources can be attributed to the different text-mining approaches: GeneRIF was curated by a mixed effort of manual annotation and various user-designed text-mining approaches, while GeneWays used natural language processing tools. The two sets of disease–gene relationships are complementary to a great extent; we, therefore, retained both resources for our web server.

Disease–gene relationships from OMIM

We downloaded the OMIM database (30), then converted annotations of the OMIM Gene Map to UMLS concepts using the UMLS natural language processing tool (26). This analysis resulted in 18 710 disease–gene relations, including 4457 diseases and 3749 genes.

After collecting the above data, we constructed a disease–disease connectivity network summarizing the molecules shared between diseases. We evaluated the connection strength between two diseases based on the enrichment of genes that are relevant to both diseases. The statistical significance of the connection between two diseases is assessed by a hypergeometric test on shared genes derived from the GWAS, OMIM, DEG, GeneRIF and GeneWays sources. We did not include disease–disease connections derived from GeneRIF and GeneWays in this server, because the disease–gene relationships extracted from the curated literature have low performance in terms of the relatedness to the comorbidity and drug treatment data (see Figures 3 and 4). In addition, those literature-based disease–gene relationships were included in the server as the complementary data.

Diseases with shared molecular mechanisms are more likely to have clinical comorbidity

We assessed the strength of disease connections based on the significance of the number of genes involved in both disorders, which were derived from genomic data and the literature. We are interested in the extent to which these mechanism-based disease connections are related to disease comorbidity relationships of the same disease pairs, based on patient clinical records. The mechanism-based and comorbidity-based disease networks can be compared in terms of the fraction of shared edges. To construct the comorbidity network, we used only high-quality comorbidity relationships, with P-values <0.01, retrieved from the PDN (1). So that the comparison is unbiased, we excluded disease pairs with ancestor–descendant relationships in the disease semantic hierarchy (i.e. ‘is-a’ relations of UMLS disease concepts). The results of this comparison are shown in Figure 3. The overall trend is that two diseases with stronger mechanism-based connections are more likely to have a significant clinical comorbidity relation. When we randomize the mechanism-based disease network by degree-preserving edge rewiring (31) , this trend disappears and the random network shares few edges with the disease comorbidity network.

Among all data sources, the disease network derived from GWAS records showed the highest level of consistency with the comorbidity network. Interestingly, the OMIM-derived network not only is less consistent but has the ‘seesaw’ performance curve. The DEG- and GWAS-based networks both have stable performance. A possible explanation is that although OMIM has high-quality data and has expanded to polygenic complex maladies in recent years (4,32), many (i.e. 35.5%) of its records are monogenic disorders for which our assessment of disease connection strength may not be precise. Finally, it is not surprising that the networks based on GeneRIF and GeneWays data from the literature have the worst performance because they include too much noisy data incurred by the use of automatic human language processing tools. However, both these networks still perform better than chance. Overall, these results validate our claim that diseases with shared molecular mechanisms are more likely to have comorbidity.

Diseases with shared molecular mechanisms are more likely to have the same drug treatment

To validate the intuitive hypothesis that diseases with shared molecular mechanisms are more likely to be treated by the same drug(s), we performed a network comparison similar to that described above. In this case, we calculated the fraction of mechanism-based disease connections for which both diseases are treated by the same drug or drugs. The drug–disease treatment resources were collected from NDFRT (see the Material and Methods section). We individually evaluated the disease networks derived from different sources (GWAS/OMIM/DEG/GeneWays/GeneRIF). For an unbiased comparison, we used only those diseases with drug–treatment relations and excluded disease pairs with ancestor–descendant relationships in the disease semantic hierarchy (i.e. ‘is-a’ relations of UMLS disease concepts).

Figure 4 displays the results of this comparison: (i) the more significant the mechanism-based connection between two diseases (derived from any data sources), the higher the likelihood they can be treated by the same drugs. For example, for the data source OMIM, we identified 1243 disease–disease connections (covering 424 diseases and 599 drugs) in each of which both diseases share significant (P-value <0.01) common molecular mechanisms and both diseases have known drug–treatment data. Among those, the fraction in which both diseases are treated by the same drug(s) for OMIM is 33%. Figure 4 shows that this fraction increases with the strength (i.e. P-value) of mechanism-based disease connections. (ii) Connections derived from the GWAS and OMIM data sources are the strongest predictors of common drug treatments. It is also interesting to note that while connections derived from text mining (GeneRIF/GeneWays) are not good predictors, they still perform better than random networks. The GWAS data source consistently outperforms OMIM by a small amount. The chance for two diseases to be treated by the same drug(s) increases with the significance of their GWAS- or OMIM-based connection; however, DEG-based connections show no such trend. This implies that the molecular mechanisms indicated by GWAS and OMIM records are more closely relevant to drug treatment than those present in the DEG records. A possible reason is that DEG records, derived from gene expression data, may more likely reflect downstream effect rather than causes of diseases.

Diseases modules in the connectivity network are more likely to have the same drug treatment

Having confirmed that disease pairs with significant mechanism-based connections are likely to be treated by the same drug(s), we further hypothesize that diseases comprising a dense subgraph in the disease connectivity network, termed a disease module (33), are likely to benefit from the same drug treatment. To validate this hypothesis, we constructed a large, high-quality disease connectivity network, comprising 5189 connections with hypergeometric P-values < 1e-6 for disease module discovery. We applied MODES, a network clustering method to discover overlapping dense clusters (34) in this network and identified 141 distinct disease modules (details see Supplementary Table S1). We ran MODES using the following parameters: minimum module size 3, maximum module size 30 and density cutoff 0.7. In 38 (27%) of the discovered modules, more than half of the member diseases can be treated by the same drug. In contrast, our randomization test (we generated 141 similar-sized modules by randomly permuting drug–treatment relations) has only 3 (2.1%) such modules. This result indicates that highly connected diseases are more likely to have the same drug treatment. Figure 5 illustrates Module-109 as an example. This densely connected disease subnetwork consists of nine diseases, of which the drug ‘fludarabine’ can treat five. Fludarabine is effective in the treatment of chronic lymphocytic leukemia (35). Although multiple myeloma (MM) and acute lymphocytic leukemia (ALL) are not currently treated with fludarabine, we found strong evidence for a potential effect on MM and ALL in the literature: (i) Caballero-Velázquez et al. reported that a combined treatment consisting of bortezomib, fludarabine, and melphalan showed advantage in high-risk MM patients (36). (ii) Daly et al. reported that fludarabine and busulfan achieved excellent outcomes with older ALL patients (37). (iii) There are several ongoing clinical trials using fludarabine to treat MM (e.g. ClinicalTrials.gov ID: NCT01131169 and NCT01453101) and ALL (e.g. ClinicalTrials.gov ID: NCT01435447 and NCT01572662). Therefore, identifying disease modules such as this one in the disease connectivity network can point the way to effective drug repositioning strategies.

Example of a disease query (arthritis)

When researching a disease, biologists usually consider what other diseases are relevant to it. In the disease view, users can enter a full or partial disease name in the input field. The server automatically matches and completes a partial name with a list of full disease names. For example, when entering ‘arthritis’, the web server yields the disease–disease network shown in the left panel of Figure 2B. In this network, a disease–disease connection indicates that two diseases involve a significant number of shared genes based on the GWAS/OMIM/DEG records. If the user clicks a disease (node) in the network, the web server shows the network of all molecules associated with this disease. For example, the right panel of Figure 2B shows the disease–gene–drug network associated with arthritis, which includes the following molecules: (i) genes associated with arthritis based on the GWAS/OMIM/DEG/GeneRIF/GeneWays records, (ii) drugs that target these genes and (iii) drugs that treat arthritis.

Given the heterogeneous diseases connected to arthritis in Figure 2B, it is worth noting that arthritis is already known to relate to large group diseases, subsuming its Diagnosis Related Group (DRG). DRG is a system to classify hospital cases into one of originally 467 groups (38) and is currently the main system documenting disease groups. DRG however has major shortcomings being highly focused and directly linked to specific molecular mechanisms in some cases (such as mitochondriopathies or muscular dystrophies) and highly nonspecific with a clinical syndrome emphasis in others (such as heart failure). Our system, however, provides a comprehensive and unbiased way to identify diseases sharing molecular similarities, given a large amount of collected genomic and literature data.

Examples of disease pair queries and their applications to revealing potential drug treatments

After viewing all diseases and molecules related to a query disease, users can further explore the shared molecular mechanism(s) between two specific diseases by clicking an edge in the disease network. For example, when a user queries MM, the web server generates a disease–disease network. The user can then click the connection between MM and hemorrhagic disorders (HD) to view a network of shared genes and drugs between HD and MM. Figure 6A shows a small part of this network with the shared gene PSMB5 and the drug bortezomib. PSMB5 is a DEG for both HD and MM, and also has a GeneRIF association with MM. Bortezomib is a therapeutic proteasome inhibitor for treating MM, and targets PSMB5 (39). Interestingly, bortezomib can also treat HD (40), suggesting that shared disease genes can serve as good targets for drug repositioning strategies.

Figure 6B shows another example network for the queried disease pair ‘arthritis’ and ‘Crohn disease’, both involve the tumor necrosis factor (TNF), based on the data sources GWAS and GeneRIF. Thalidomide is an immunomodulatory drug that targets TNF. Interestingly, thalidomide can treat both arthritis and Crohn diseases (41), supporting our hypothesis that diseases with a shared molecular mechanism are likely to be treated by the same drug. Another drug shown in Figure 6B, ‘adalimumab’, is known to target TNF and can treat arthritis, suggesting that it may also be effective for Crohn disease. In fact, this inference is confirmed independently by a recent report of Peters et al. (42). Also, several clinical trials about Adalimumab and Crohn disease (e.g. ClinicalTrials.gov ID: NCT01556672 and NCT01958827) are currently ongoing.