Anna Bauer-Mehren; Markus Bundschus; Michael Rautschka; Miguel A. Mayer; Ferran Sanz; Laura I. Furlong

“ OMIM has become a highly popular source in medical genetics [6] . It traditionally focused on mendelian diseases and later started to include complex diseases as well. In the last years, other databases have been built, among them the Pharmacogenomics Knowledge Base (PHARMGKB) specialized on the knowledge about genes that are involved in modulating drug response [7] or the Comparative Toxicogenomics Database (CTD) focused on the effect of environmental chemicals on human disease [8] . Each of the databases focuses on different aspects of phenotype-genotype relationships. ”

“ On the other hand, for diseases with more than 10 gene products annotated, it is more likely that more than one pathway is involved. Moreover, it is striking that although the text-mining derived network is very dense with an average of 18.7 genes per disease, the homogeneity values still differ significantly from random, with values comparable to the ones in CURATED. The observed decrease of homogeneity was observed for all databases including OMIM suggesting that it is not related to the integration effort but arises due to increased locus heterogeneity of diseases. ”

“ The HINscore is 1 if all nodes in a cluster are directly connected in HIN; it is 0 if they are not connected in HIN but only in the gene cluster due to shared disease associations. Therefore, the HINscore represents a means to measure to what extent the gene clusters represent functional modules and hence to measure the modularity of human diseases. ”

“ Diseases are coloured according to their disease class (see Fig. S4) . The most significantly enriched Jak-STAT signaling pathway is displayed with some nodes from the cluster colored in red (right part) . GO term and pathway enrichment analysis showed that for this heterogeneous cluster, there are common biological processes associated with the distinct diseases. For instance, although there are 59 pathways annotated to this cluster, the pathway homogeneity is 41% indicating that almost half of the gene products appear in the same pathway. ”

“ In this section, we assess the modularity from the gene centric view by evaluating if groups of phenotypically related genes represent functional modules in the cell. For this purpose we analyzed the functional annotations of gene products in clusters derived from applying the graph-clustering algorithm on the gene projection networks (see Methods) .
Gene products can be functionally related to each other in different ways: by means of direct, physical protein-protein interactions or by more indirect associations as observed between enzymes in the context of a metabolic pathway. ”

“ Our computational analysis has important implications for understanding disease mechanisms and how environmental factors, such as drugs, influence human health. Finally, we provide all gene-disease networks in a user-friendly way through DisGeNET [28] , a Cytoscape plugin and as SQLite database for direct computational access to aid future research of disease-related processes that will also benefit drug discovery and development. ”

“ The literature-derived human gene-disease network (LHGDN) is a text mining derived database with focus on extracting and classifying gene-disease associations with respect to several biomolecular conditions. It uses a machine learning based algorithm to extract semantic gene-disease relations from a textual source of interest. ”

“ The disease classification allows filtering and searching of the network restricted to disease class, all implemented within DisGeNET [28] . The disease classification is available as text file in the supplementary material (Text S2) . Many diseases are assigned to more than one disease class as several systems or organs are affected. Fig. S4 shows the disease color mapping used in DisGeNET [28] . Disease and gene nodes can be colored according to their disease class and can have multiple colors if they are assigned to more than one disease class. ”

“ Online Mendelian Inheritance in Man (OMIM) focuses on inherited diseases. Gene-disease associations were obtained by parsing the mim2gene file for associations of type “phenotype” (data was downloaded from ftp://ftp.ncbi.nlm.nih.gov/gene/DATA/mim2gene on June, 6th 2009) and classified as Marker in our gene-disease association ontology (see Fig. S1) . In total, we obtained for 2198 distinct genes and 2473 distinct disease terms 3432 gene-disease associations. After mapping of disease vocabularies, the OMIM network contained 2417 distinct diseases. ”

“ All in all, cluster analysis of the gene projection network uncovered an interesting relationship between environmental exposure to arsenic compounds and breast cancer. This relationship deserves further investigation at the epidemiological and molecular levels.
Case study 4: Identification of shared mechanisms of different diseases
Another cluster containing 79 genes is an example of a heterogeneous cluster in which genes are associated with different diseases. ”

“ Moreover, influences of environmental factors, infectious agents or drugs have to be considered when studying the occurrence and evolution of a disease. In complex diseases, alterations in several genes can make subtle contributions to the susceptibility of a particular individual. ”

“ The projected networks contain only vertices of the same kind (monopartite) and two nodes are connected if they share a neighbor in the original bipartite graph. Before calculating node degree distributions and projecting the networks, we simplified the graphs by removing multiple edges. In the simplified graphs the node degree represents the number of first neighbors.
All gene-disease networks are available as SQLite database and through DisGeNET, a Cytoscape plugin, both available online at http://ibi.imim.es/DisGeNET/DisGeNETweb.html#Download. ”

“ It could be argued that such genes encode multifunctional proteins that participate in different biological processes, and mutations in these proteins affecting different functions can then lead to different disease phenotypes. ”

“ In conclusion, clustering analysis of the gene projection network followed by functional enrichment analysis can be used to propose new candidate disease genes. We would also like to mention here that the CURATED set of DisGeNET represents a suitable dataset to benchmark computational methods for the prediction of candidate disease genes. ”

“ On the basis of this knowledge it is possible to create a hypothesis on the mechanisms by which certain drugs such as Perhexiline can lead to Rhabdomyolysis. Perhexiline, which is prescribed for severe Angina Pectoris [44] inhibits CPT1, shifting myocardial substrate utilization from long chain fatty acids to carbohydrates (http://www.drugbank.ca/drugs/DB01074) . Perhexiline can also target, to a lesser extent, CPT2 [45] , which would explain the toxic effects of the drug in skeletal muscles due to the association of CPT2 with Rhabdomyolysis. ”

“ Pathway homogeneity for individual diseases.Mean pathway homogeneity values of single diseases and random controls are plotted for all four networks binned by the number of associated gene products per disease. ”

“ The latter are neurometabolic disorders that result from defects in the mitochondrial respiratory chain. Genes associated with these diseases encode proteins that form a physically interacting module as illustrated by the HIN subgraph. Other examples of clusters with high HINscore are related to peroxisomal disorders (e.g. Zellweger syndrome) , different types of anemia (Diamond-Blackfan anemia or Heinz body anemia) or Walker-Warburg and Fukuyama syndromes. ”