ABSTRACT
Historically, discovery of new therapeutic drugs has relied on trial-and error-approaches, natural products, or even serendipitous findings (the discovery of penicillin being one the most famous examples). The field rapidly expanded in the past century with the development of modern chemistry and the industrialization of the drug discovery process. However, the process of developing a new drug is costly, very long, and most often unsuccessful. Indeed, the pace of discovery of new therapeutic drugs has been consistently decreasing for the last years, and most of the current drugs are acting on a small number of biological targets [45]. This is particularly contradictory in a time in which technical advances have helped characterize complete genomes, including human ones, and have made available a humongous amount of information potentially useful for drug discovery. In this context, existing computational methods for target identification and characterization, ligand binding prediction (docking), and virtual screening aim to expand the current set of available targets and boost the discovery of new active compounds. Nevertheless, the application of these methods usually comes at a high computational cost. For instance, average computational times for virtual screening with state-of-theart DOCK have been reported to be around one ligand per minute on a Xeon 3.0 GHz processor [47]. Therefore, a standard small-compound library, like the ZINC database [32] with over 13 million purchasable compounds (as of October 2009), would need a minimum of 90 days on 100 Xeon 3.0 GHz processors to be fully screened. With more computational power, provided by parallel architectures like cell broadband engine (Cell BE), more conformations per compound could be considered, thus expanding the conformational space of the chemical compounds. Furthermore, the size of the libraries could be enlarged, aiming to better cover the chemical space.
