NETWORK INFERENCE DRIVEN DRUG DISCOVERY

The application of rational drug design principles in the era of network-pharmacology requires the investigation of drug-target and target-target interactions in order to design new drugs. The presented research was aimed at developing novel computational methods that enable the efficient analysis of complex biomedical data and to promote the hypothesis generation in the context of translational research. The three chapters of the Dissertation relate to various segments of drug discovery and development process. The first chapter introduces the integrated predictive drug discovery platform „SmartGraph”. The novel collaborative-filtering based algorithm „Target Based Recommender (TBR)” was developed in the framework of this project and was validated on a set of 28,270 experimentally determined bioactivity data points involving 1,882 compounds and 869 targets. The TBR is integrated into the SmartGraph platform. The graphical interface of SmartGraph enables data analysis and hypothesis generation even for investigators without substantial bioinformatics knowledge. The platform can be utilized in the context of target identification, drug-target prediction and drug repurposing. The second chapter of the Dissertation introduces an information theory inspired dynamic network model and the novel “Luminosity Diffusion (LD)” algorithm. The model can be utilized to prioritize protein targets for drug discovery purposes on the basis of available information and the importance of the targets. The importance of targets is accounted for in the information flow simulation process and is derived merely from network topology. The LD algorithm was validated on 8,010 relations of 794 proteins extracted from the Target Central Resource Database developed in the framework of the “Illuminating the Druggable Genome” project. The last chapter discusses a fundamental problem pertaining to the generation of similarity network of molecules and their clustering. The network generation process relies on the selection of a similarity threshold. The presented work introduces a network topology based systematic solution for selecting this threshold so that the likelihood of a reasonable clustering can be increased. Furthermore, the work proposes a solution for generating so-called “pseudo-reference clustering” for large molecular data sets for performance evaluation purposes. The results of this chapter are applicable in the lead identification and development processes