The identification of disease genes from candidated regions is one of the most important tasks in bioinformatics research. Most approaches based on function annotations cannot be used to identify genes for diseases without any known pathogenic genes or related function annotations. The authors have built a new web tool, DGHunter, to predict genes associated with these diseases which lack detailed function annotations. Its performance was tested with a set of 1506 genes involved in 1147 disease phenotypes derived from the morbid map table in the OMIM database. The results show that, on average, the target gene was in the top 13.60% of the ranked lists of candidates, and the target gene was in the top 5% with a 40.70% chance. DGHunter can identify disease genes effectively for those diseases lacking sufficient function annotations.
Top1. Introduction
The identification of genes involved in inherited human diseases plays an important role in elucidating pathogenesis and developing diagnosis and prevention measures. Through complex-trait linkage studies, many disease genes are located within one or more specific chromosomal regions (McCarthy, Smedley, & Hide, 2003). It is time-consuming and labor-intensive to perform random mutation analysis for the hundreds of genes in the regions of interest. Clearly, predicting the best candidate genes by computational approaches is necessary to facilitate the identification of disease-related genes for further study.
Disease phenotypes provide a window into the gene function. Franke et al. (2006) observed that GO annotation is the most effective data resource, and the accuracy based on GO was slightly improved by adding other types of data. Several approaches to identify disease related genes based on function annotations have been presented in recent years (Perez-Iratxeta, Bork, & Andrade, 2002; Perez-Iratxeta, Wjst, Bork, & Andrade, 2005; Perez-Iratxeta, Bork, & Andrade-Navarro, 2007; Turner, Clutterbuck, & Semple, 2003). However,these tools often suffer from annotation bias as they cannot deal with diseases lacking known causative genes. Neither can these tools handle known genes lacking sufficiently detailed function annotations. The phenotype method of G2D developed by Perez-Iratxeta et al. can be used to make prediction for those diseases (Perez-Iratxeta, Bork, & Andrade, 2002; Perez-Iratxeta, Wjst, Bork, & Andrade, 2005). It firstly mines the disease-related GO terms from the MEDLINE/PubMed database, then associates the RefSeq sequence with the disease according to their function annotations, finally prioritizes the candidate genes by sequence similarity searches. The assumptions of this method are that similar sequences are paralogs, and that paralogs have the same or similar function. However, function similarity does not always require sequence similarity. For example, both P53 (Vogelstein & Kinzler, 1992) and BRCA1 (Thompson, Jensen, Obermiller, Page, & Holt, 1995) function as tumor suppressor genes. Similar to BRCA1, mutations in P53 have also been found in breast cancer patients (Davidoff, Humphrey, Iglehart, & Marks, 1991). However, the two genes share no sequence homology. As a result, sequence similarity may not be able to reveal functional similarity. Therefore, the assumptions could easily produce false-positive and false-negative results.
To avoid this problem, we have developed a computational web tool, DGHunter. It allows users to predict candidated causative genes for a genetic disease lacking known causative genes. Starting with the assumption that genes involved in phenotypically similar diseases share similarity in their functional annotation (Jimenez-Sanchez, Barton, & David, 2001) DGHunter makes prediction by using a combination of text mining and gene-function similarity analysis. The process of the text mining of DGHunter adopts similar approach as that of G2D, and the function-similarity analysis uses a novel algorithm based on the GO directed acyclic graphs (DAG). More details can be found in the following sections.