Yixuan Chen; Wenhui Wang; Yingyao Zhou; Robert Shields; Sumit K. Chanda; Robert C. Elston; Jing Li

“ For many data sources, one has to measure the relationships between candidate genes and disease genes directly. For other data sources, such as PPI networks, one can either choose to measure the gene-gene relationships locally, or measure them globally. Köhler et al. [11] have shown that global measures perform better than local measures for prioritizing disease genes using PPI networks. A fundamental issue in studies using a single data source is the potential bias of their results caused by the incompleteness and noise of one particular data set. ”

“ However, great challenges exist because many common diseases are caused by multiple disease genes with small to moderate effects. Even diseases that show Mendelian inheritance may involve multiple genes due to heterogeneity. Gene-gene interactions, as well as gene-environment interactions, also play an important role in the development of diseases. Classifications of diseases, which are mostly based on observed phenotypes, may not necessarily reflect their underlying mechanisms. ”

“ We selected three disease families as an example to show the improvement by approaches using multiple data sources (Table 2) . The informativeness of networks on all diseases can be found in Dataset S1. In one example, the disease family “Generalized epilepsy with febrile seizures plus” obtains little information from the PPI network. Therefore it was not surprising that the RWR, which depends on the PPI network solely, could not correctly predict disease genes in the cross-validation. In contrast, the gene co-expression network provided sufficient information about their connections. ”

“ For a specific disease, the prediction result will be limited by existing knowledge about the disease, including the number of known disease genes and their relationships within the existing data sources. We have used the concept of disease families in order to increase the number of known disease genes in each family. Some recent studies have considered relationships/similarities between diseases/phenotypes [29] and have utilized phenotype similarities in their gene prioritization approach [30] – [31] . We will investigate approaches to incorporate phenotype similarities into our framework. ”

“ Gene-gene relationships and gene-disease relationships are then defined based on the overall topology of each network using a diffusion kernel measure. These relationship measures are in turn normalized to derive an overall measure across all networks, which is utilized to rank all candidate genes. Based on the informativeness of available data sources with respect to each specific disease, we also propose an adaptive threshold score to select a small subset of candidate genes for further validation studies. ”

“ Abstract Identifying disease genes is crucial to the understanding of disease pathogenesis, and to the improvement of disease diagnosis and treatment. In recent years, many researchers have proposed approaches to prioritize candidate genes by considering the relationship of candidate genes and existing known disease genes, reflected in other data sources. In this paper, we propose an expandable framework for gene prioritization that can integrate multiple heterogeneous data sources by taking advantage of a unified graphic representation. ”

“ A variety of technologies exist to measure the levels of many such activities. Over the years, a vast amount of data from different sources has been accumulated and stored in a huge number of biological databases, many of which are publicly available. For a particular disease, such as breast cancer, tissue gene expression data might exist in some databases. Known disease genes and their interacting parters may have been recorded in protein-protein interaction (PPI) databases. Researchers may have also collected and constructed disease pathways based on previous studies. ”

“ Even though the networks are quite comprehensive, some disease genes may not occur in a network at all, or may have limited connections. Therefore, for some disease families, it is not appropriate to use the data sources to prioritize genes if the networks themselves do not contain enough information about these disease families. ”

“ Missense mutation in both alleles of the TH gene is known to cause dopamine-related phenotypes, including dystonia and infantile Parkinsonism. Most recently, a study has found a rare novel deletion of the entire TH gene in an adult with PD [12] . The result from this study had not been entered into the OMIM database. This clearly shows the value of our in silico prioritization approach, and the top ranked genes returned by our approach should receive more attentions in follow-up or validation studies. We have also tested RWR and ENDEAVOUR on the same data set. ”

“ In addition, researchers have increasingly realized that there are many levels of controls along the paths from genotypes to phenotypes, resulting in a weaker relationship between genotypes and phenotypes [1] that may or may not be captured using traditional linkage or association approaches. Furthermore, linkage analysis usually can only identify chromosomal intervals that may contain up to hundreds of candidate genes owning to the limited number of crossovers in sampled families. Genome-wide association studies may also return many regions that show moderate to high signals. ”

“ If a candidate gene is closer to the disease genes than the disease genes are to themselves on average, this candidate gene is more likely to be associated with the disease, too. This meta score can be used as a threshold for declaring significant candidate genes. In the Results section, we will discuss the use of File:Pone.0021137.e107.jpg and its variants as “adaptive ranking thresholds” and evaluate their performance in comparison with the top-File:Pone.0021137.e108.jpg approach.
Informativeness of a Network
The informativeness of networks is different for different disease families. ”

“ The final protein-protein interaction network contains File:Pone.0021137.e001.jpg human genes that encode proteins in the network and File:Pone.0021137.e002.jpg edges. This exact dataset has been used in other previous biological studies [18] , [19] .
Human Gene Expression Data
The human tissue expression dataset was obtained from GNF's SymAtlas web site [20] . This dataset consists of 79 human tissues in duplicates, measured using the Affymetrix U133A array that consists of 22,215 probe sets. All array measurements were processed and normalized using the Affymetrix MAS5 algorithm. ”

“ Second, we also use the measure of the receiver operating characteristic (ROC) curve, which shows the relation between the sensitivity (true positive) and the specificity (true negative rate) by varying the threshold for declaring positives. ”

“ File:Pone.0021137.e034.jpg, where File:Pone.0021137.e035.jpg is a tuning parameter and File:Pone.0021137.e036.jpg, File:Pone.0021137.e037.jpg being a diagonal matrix with the diagonal elements containing the node degrees. The diffusion kernel represents a global similarity between nodes in a graph, with higher values representing closer relationships. ”