Estratégias de otimização e paralelização massiva do programa de atracamento molecular DockThor

• Main Author: Leão, Aaron Bruno
• Publication Date: 2022
• Format: Doctoral thesis
• Language: por
• Source: Biblioteca Digital de Teses e Dissertações do LNCC
• Download full: https://tede.lncc.br/handle/tede/354
• topic: Atracamento molecular; Molecular docking; Algorítmos genéticos; Bioinformática; DockThor (Programa de computador); CNPQ::CIENCIAS BIOLOGICAS::BIOLOGIA GERAL

The molecular docking methodology is a significant tool in Structure Based Drug Design, which aims to find the binding mode of a small molecule with a receptor as well as its binding affinity. This technique helps to reduce the cost, time and number of failures in the development of new drugs. The DockThor molecular docking program developed in the Molecular Modeling of Biological Systems Group at LNCC, hosted at <dockthor. lncc.br> and implemented in the Santos Dumont supercomputer serves the national and international scientific community and its number of accesses grows every year. Even though DockThor program presents competitive performance with state-of-the-art programs, DockThor is not competitive in terms of execution time, which makes it difficult and expensive to use in virtual screening experiments containing millions of compounds. In this work, we developed the adaptation of the grid calculation step so that it could be used more efficiently in GPU devices. This improvement enabled a significant performance increase (20 times), which computationally enables the implementation of several ensemble docking strategies and the composition of multiple grids. The structures involved in the steps of the genetic algorithm of multiples mimima were also adapted using phenotypic crowding. A version of the steady-state algorithm was implemented with the same energy accuracy and performance gain of more than 30% in the evolution. This result implies a better performance in virtual screening experiments. Aiming at obtaining computational performance in GPU-like architectures, a new generational algorithm was developed that was capable of performing the independent evolution steps and with the characteristic of being individually parallelized. This new algorithm, which also takes advantage of coalescing structures, performed well in the energetic analysis of the test set and prediction of the native pose of receptor-ligand complexes. Finally, a new generational algorithm was designed in OpenCL (from the generational produced and tested) that performed all stages of evolution on the GPU, with no data communication with the CPU during the evolution of the algorithm. This generational GPU algorithm achieved the energy accuracies of the original steady-state and still achieving a performance of 3.9 to 7.3 times faster than the original algorithm that performed all stages of evolution on the GPU, with no data communication with the CPU during the evolution of the algorithm. This algorithm achieved the energy accuracies of the original steady-state and still achieving a performance of 3.9 to 7.3 times faster than the original algorithm. The results obtained in this work point to the feasibility of using the DockThor program in virtual screenings containing millions of compounds.