Work

Network Biology for Gene Prioritization

Year: 2011
Type: dissertation
Abstract: Motivation:Genetic studies and high-throughput genome wide screens can identify genes and proteins that are candidate members for a biological process of interest (such as disease and pathways). The disadvantage of the screens is their identification of tens or hundreds of candidate genes. Aerts et al. (2006) developed gene prioritization methods that rank candidate genes based on their similarity to genes already associated with a disease or process, using multiple data sources (such as sequence, expression, literature, etc), ENDEAVOUR. However, this method cannot prioritize candidates if no similar genes can be identified a priori. This prevents the method from tackling truly innovative discoveries, when little is known about a disease. Currently, there are no well-established gene prioritization strategies without a set of training genes. The new strategy I am working on, consists of checking not only the expression of a candidate, but also of its "partner" genes in a gene network derived from multiple sources. A strong candidate should have many partners that are differentially expressed, meaning that it belongs to a disrupted expression module. Project description:A mouse model network based on the STRING database (http://string.embl.de/) was built, for which the similarities between the single genes could be computed. Candidate genes with genes in their neighborhood having highly differentially expressed levels are strong candidates for causing a disease.Different kernel matrices were used as global similarity measures, e.g. the Exponential diffusion Kernel (Konder et al. 2002), the Commute time Kernel (Fouss et al. 2006), the Von Neumann Kernel (Kandola et al. 2003). The aim of these kernel matrices is to consider global similarity instead of local similarity in the network. They are able to detect similarities between genes that have no connection in the STRING network, although there are indirect connections via other genes. The resulting global similarity network is densely connected and contains all direct and indirect connections between single genes in the network, whereby the value of the connections represents the degree of similarity of two single genes.To improve the enormous computing time, the kernel matrices were approximated by the complete or incomplete Cholesky decomposition and the Reduced Eigenvalue decomposition. Although the result suffers from a quality loss depending on a user defined threshold, the computation receives large gains in CPU time.Knowing the similarities between single genes in the network, the differential expression level of the genes themselves, as well as the differential expression level of adjacent genes can be considered, based on microarray experiments. Considering a specific disease, a microarray experiment is performed and the differential expression levels of all genes in the network are computed. A list of potential candidate genes is required, which can be derived from publications, OMIM etc. For each candidate gene, the differential expression levels of all adjacent genes are ordered by their similarity and ranked. The higher the variability in their differential expression levels for similar genes, the better the candidate gene and the higher its final ranking score. For computing the score the first implementations are realized, although they need further improvement.The microarray experiments that are used until now, are published datasets with known lists of candidate genes. In the future this method will be applied to congenital heart defects studies ongoing at Leuven University Hospitals.The final aim is to develop a software application for the computational prioritization of candidates genes, based on microarray experiments.Another goal is to find new potential candidate genes after ranking the whole genome. Although genes that do not belong to the list of known candidate genes can have high ranking scores and might play a role in the disease studied in the microarray experiments. These genes might be new candidate genes that can be part of new clinical studies.
Source: Lirias
Institution: KU Leuven
Language: English
Open Access status: green
OpenAlex ID: W7021100535
Cites: 240
Cited by:
Related to: 10