Docktdeep: um programa para desenvolvimento de funções scoring para triagem virtual de compostos usando redes neurais profundas

• Ano de defesa: 2020
• Autor(a) principal: Silva, Matheus Müller Pereira da
• Orientador(a): Dardenne, Laurent Emmanuel
• Banca de defesa: Dardenne, Laurent Emmanuel, Custódio, Fábio Lima, Sant'Anna, Carlos Mauricio Rabello de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Laboratório Nacional de Computação Científica
• Programa de Pós-Graduação: Programa de Pós-Graduação em Modelagem Computacional
• Departamento: Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA)
• País: Brasil
• Palavras-chave em Português: DockTDeep (Programa de computador); Planejamento de fármacos; Funções scoring
• Área do conhecimento CNPq: CNPQ::CIENCIAS BIOLOGICAS::BIOQUIMICA
• Link de acesso: https://tede.lncc.br/handle/tede/323

Molecular docking and virtual screening (VS) methodologies are crucial to modern drug design. As a fundamental part of this methodologies are the scoring functions (SF), which are used to estimate the binding affinity between a molecule (ligand) and a pharmacological target of interest, usually a protein. While determination of the correct three-dimensional geometry of ligands in active sites of targets is achieved with relatively satisfactory accuracy by several molecular docking software, the accurate prediction of protein-ligand binding affinity remains a challenge. SFs capable of satisfactorily predicting the binding affinity for several different targets and ligands are essential to a successful VS methodology. In past years, the employment of machine learning techniques in the development of SFs has been gaining prominence, overcoming the performance of more traditional approaches based on physical principles. In this work, a computer program called DockTDeep was implemented to support the development and application of SFs based on deep learning (DL) techniques. The software is capable of generating different molecular representations of protein-ligand complexes, based on multidimensional grids of voxels (4D tensors). Another crucial feature of the program is to enable the management of large data sets of protein-ligand complexes, appropriately for training with DL algorithms. Besides, in this work, SFs were also developed using the DockTDeep software to validate its functioning. SFs developed in this work were trained using data from the PDBbind refined set and tested in the PDBbind core set v.2013 and v.2016 validation sets. Best Pearson correlations for the predictions in each of the validation sets were R = 0,645 and R = 0,685, respectively. The DockTDeep software proved to suitable for the development of SFs based in DL, which uses voxel representation of molecules. Several models and representations can be developed using this software, and models obtained are well suited to be used in large-scale VS experiments.