Abstract

In the current drug discovery process, the identification of new target proteins and potential ligands is very tedious, expensive and time-consuming. Thus, use of in silico techniques is of utmost importance and proved to be a valuable strategy in detecting complex structural and bioactivity relationships. Increased demands of computational power for tremendous calculations in scientific fields and timely analysis of generated piles of data require innovative strategies for efficient utilization of distributed computing resources in the form of computational grids. Such grids add a new aspect to the emerging information technology paradigm by providing and coordinating the heterogeneous resources such as various organizations, people, computing, storage and networking facilities as well as data, knowledge, software and workflows. The aim of this study was to develop a university-wide applicable grid infrastructure, UVieCo (University of Vienna Condor pool) which can be used for implementation of standard structure- and ligand-based drug discovery applications using freely available academic software. Firewall and security issues were resolved with a virtual private network setup whereas virtualization of computer hardware was done using the CoLinux concept in a way to run Linux-executable jobs inside Windows machines. The effectiveness of the grid was assessed by performance measurement experiments using sequential and parallel tasks. Subsequently, the association of expression/sensitivity profiles of ABC transporters with activity profiles of anticancer compounds was analyzed by mining the data from NCI (National Cancer Institute). The datasets generated in this analysis were utilized with ligand-based computational methods such as shape similarity and classification algorithms to identify and separate P-gp substrates from non-substrates. While developing predictive classification models, the problem of imbalanced class distribution was proficiently addressed using the cost-sensitive bagging approach. Applicability domain experiment revealed that our model not only predicts NCI compounds well, but it can also be applied to drug-like molecules. The developed models were relatively simple but precise enough to be applicable for virtual screening of large chemical libraries for the early identification of P-gp substrates which can potentially be useful to remove compounds of poor ADMET properties in an early phase of drug discovery. Additionally, shape-similarity and self-organizing maps techniques were used to screen in-house as well as a large vendor database for identification of novel selective serotonin reuptake inhibitor (SSRI) like compounds to induce apoptosis. The retrieved hits possess novel chemical scaffolds and can be considered as a starting point for lead optimization studies. The work described in this thesis will be useful to create distributed computing environment using available resources within an organization and can be applied to various applications such as efficient handling of imbalanced data classification problems or multistep virtual screening approach.