Data driven approaches to improve the drug discovery process : a virtual screening quest in drug discovery

Abstract

Drug discovery has witnessed an increase in the application of in silico methods to complement existing in vitro and in vivo experiments, in an attempt to 'fail fast' and reduce the high attrition rates of clinical phases. Computer algorithms have been successfully employed for many tasks including biological target selection, hit identification, lead optimization, binding affinity determination, ADME and toxicity prediction, side-effect prediction, drug repurposing, and, in general, to direct experimental work. This thesis describes a multifaceted approach to virtual screening, to computationally identify small-molecule inhibitors against a biological target of interest. Conformer generation is a critical step in all virtual screening methods that make use of atomic 3D data. We therefore analysed the ability of computational tools to reproduce high quality, experimentally resolved conformations of organic small-molecules. We selected the best performing method (RDKit), and developed a protocol that generates a non-redundant conformer ensemble which tends to contain low-energy structures close to those experimentally observed. We then outline the steps we took to build a multi-million, small-molecule database (including molecule standardization and efficient exact, substructure and similarity searching capabilities), for use in our virtual screening experiments. We generated conformers and descriptors for the molecules in the database. We tagged a subset of the database as `drug-like' and clustered this to provide a reduced, diverse set of molecules for use in more computationally-intensive virtual screening protocols. We next describe a novel virtual screening method we developed, called Ligity, that makes use of known protein-ligand holo structures as queries to search the small-molecule database for putative actives. Ligity has been validated against targets from the DUD-E dataset, and has shown, on average, better performance than other 3D methods. We also show that performance improved when we fused the results from multiple input structures. This bodes well for Ligity's future use, especially when considering that protein structure databases such as the Protein Data Bank are growing exponentially every year. Lastly, we describe the fruitful application of structure-based and ligand-based virtual screening methods to Plasmodium falciparum Subtilisin-like Protease 1 (PfSUB1), an important drug target in the human stages of the life-cycle of the malaria parasite. Our ligand-based virtual screening study resulted in the discovery of novel PfSUB1 inhibitors. Further lead optimization of these compounds, to improve binding affinity in the nanomolar range, may promote them as drug candidates. In this thesis we postulate that the accuracy of computational tools in drug discovery may be enhanced to take advantage of the exponential increase of experimental data and the availability of cheaper computational power such as cloud computing.