Algorithmic and Technical Improvements for Next Generation Drug Design Software Tools

Abstract

[eng] The pharmaceutical industry is actively looking for new ways of boosting the efficiency and effectiveness of their R&D programmes. The extensive use of computational modeling tools in the drug discovery pipeline (DDP) is having a positive impact on research performance, since in silico experiments are usually faster and cheaper that their real counterparts. The lead identification step is a very sensitive point in the DDP. In this context, Virtual high-throughput screening techniques (VHTS) work as a filtering mecha-nism that benefits the following stages by reducing the number of compounds to be tested experimentally. Unfortunately the simplifications applied in the VHTS docking software make them prone generate false positives and negatives. These errors spread across the rest of the DDP stages, and have a negative impact in terms of financial and time costs. In the Electronic and Atomic Protein Modelling group (Barcelona Supercomputing Center, Life Sciences department), we have developed the Protein Energy Landscape Exploration (PELE) software. PELE has demonstrated to be a good alternative to explore the conformational space of proteins and perform ligand-protein docking simulations. In this thesis we discuss how to turn PELE into a faster and more efficient tool by improving its technical and algorithmic features, so that it can be eventually used in VHTS protocols. Besides, we have addressed the difficulties of analyzing extensive data associated with massive simulation production. First, we have rewritten the software using C++ and modern software engineering techniques. As a consequence, our code base is now well organized and tested. PELE has become a piece of software which is easier to modify, understand, and extend. It is also more robust and reliable. The rewriting the code has helped us to overcome some of its previous technical limitations, such as the restrictions on the size of the systems. Also, it has allowed us to extend PELE with new solvent models, force fields, and types of biomolecules. Moreover, the rewriting has make it possible to adapt the code in order to take advantage of new parallel architectures and accelerators obtaining promising speedup results. Second, we have improved the way PELE handles protein flexibility by im-plemented and internal coordinate Normal Mode Analysis (icNMA) method. This method is able to produce more energy favorable perturbations than the current Anisotropic Network Model (ANM) based strategy. This has allowed us to eliminate the unneeded relaxation phase of PELE. As a consequence, the overall computational performance of the sampling is significantly improved (-5-7x). The new internal coordinates-based methodology is able to capture the flexibility of the backbone better than the old method and is in closer agreement to molecular dynamics than the ANM-based method.