%0 Generic
%A Suratanee, Apichat
%D 2012
%F heidok:13845
%K Bioinformatics, Host factor identification, Protein-Protein interaction, Machine learning, RNAi data
%R 10.11588/heidok.00013845
%T Computational Analysis of RNAi Screening Data to Identify Host Factors Involved in Viral Infection and to Characterize Protein-Protein Interactions
%U https://archiv.ub.uni-heidelberg.de/volltextserver/13845/
%X The study of gene functions in a variety of different treatments, cell lines and organisms has been facilitated by RNA interference (RNAi) technology that tracks the phenotype of cells after silencing of particular genes. In this thesis, I describe two computational approaches developed to analyze the image data from two different RNAi screens. Firstly, I developed an alternative approach to detect host factors (human proteins) that support virus growth and replication of cells infected with the Hepatitis C virus (HCV). To identify the human proteins that are crucial for the efficiency of viral infection, several RNAi experiments of viral-infected cells have been conducted. However, the target lists from different laboratories have shown only little overlap. This inconsistency might be caused not only by experimental discrepancies, but also by not fully explored possibilities of the data analysis. Observing only viral intensity readouts from the experiments might be insufficient. In this project, I describe our computational development as a new alternative approach to improve the reliability for the host factor identification. Our approach is based on characterizing the clustering of infected cells. The idea is that viral infection is spread by cell-cell contacts, or at least advantaged by the vicinity of cells. Therefore, clustering of the HCV infected cells is observed during spreading of the infection. We developed a clustering detection method basing on a distance-based point pattern analysis (K-function) to identify knockdown genes in which the clusters of HCV infected cells were reduced. The approach could significantly separate between positive and negative controls and found good correlations between the clustering score and intensity readouts from the experimental screens. In comparison to another clustering algorithm, the K-function method was superior to Quadrat analysis method. Statistical normalization approaches were exploited to identify protein targets from our clustering-based approach and the experimental screens. Integrating results from our clustering method, intensity readout analysis and secondary screen, we finally identified five promising host factors that are suitable candidate targets for drug therapy. Secondly, a machine learning based approach was developed to characterize protein-protein interactions (PPIs) in a signaling network. The characterization of each PPI is fundamental to our understanding of the complex signaling system of a human cell. Experiments for PPI identification, such as yeast two-hybrid and FRET analysis, are resource-intensive, and, therefore, computational approaches for analysing large-scale RNAi knockdown screens have become an important pursuit of inferring the functional similarities from the phenotypic similarities of the down-regulated proteins. However, these methods did not provide a more detailed characterization of the PPIs. In this project, I developed a new computational approach that is based on a machine learning technique which employs the mitotic phenotypes of an RNAi screen. It enables the identification of the nature of a PPI, i.e., if it is of rather activating or inhibiting nature. We established a systematic classification using Support Vector Machines (SVMs) that was based on the phenotypic descriptors and used it to classify the interactions that activate or inhibit signal transduction. The machines yielded promising results with good performance when integrating different sets of published descriptors and our own developed descriptors calculated from fractions of specific phenotypes, linear classification of phenotypes, and phenotypic distance to distinct proteins. A comprehensive model generated from the machines was used for further predictions. We investigated the nature of pairs of interacting proteins and generated a consistency score that enhanced the precisions of the classification results. We predicted the activating/inhibiting nature for 214 PPIs with high confidence in signaling pathways and enabled to identify a new subgroup of chemokine receptors. These findings might facilitate an enhanced understanding of the cellular mechanisms during inflammation and immunologic responses. In summary, two computational approaches were developed to analyze the image data of the different RNAi screens: 1) a clustering-based approach was used to identify the host factors that are crucial for HCV infection; and 2) a machine learning-based approach with various descriptors was employed to characterize PPI activities. The results from the host factor analysis revealed novel target proteins that are involved in the spread of the HCV. In addition, the results of the characterization of the PPIs lead to a better understanding of the signaling pathways. The two large-scale RNAi data were successfully analyzed by our established approaches to obtain new insights into virus biology and cellular signaling.