Deep learning para seleção de conformações de proteínas considerando suas propriedades tridimensionais

• Ano de defesa: 2018
• Autor(a) principal: Silva, Raphael Giordano do Nascimento e
• Orientador(a): Winck, Ana Trindade
• Banca de defesa: Machado, Karina dos Santos , Silva, Luís Alvaro de Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Centro de Tecnologia
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
• Departamento: Ciência da Computação
• País: Brasil
• Palavras-chave em Português: Aprendizagem profunda; Redes neurais convolucionais; Docagem molecular
• Palavras-chave em Inglês: Deep learning; Convolutional neural networks; Molecular docking
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.ufsm.br/handle/1/16579

A new drug discovery is financially costly and time-consuming. To deal this issue, the Rational Drug Design process investigates the macromolecular interaction between receptor molecules and drug-candidate ligands. Through molecular docking experiments it becomes possible to evaluate the binding quality between these molecules. To simulate the flexibility of the receptor, molecular dynamics simulations can be performed, where the protein is represented by a distinct conformation at each instant of time. Thus, molecular docking experiments can be performed on these different conformations. Several machine learning techniques can be used to mine these data. This work presents an approach for artificial neural networks deep learning, which uses as input several conformation about a given protein, generated through molecular dynamic simulations. These conformation are described in terms of its atoms tridimensional coordinates, labeled by a target attribute FEB, which points out the quality about each conformation in molecular docking experiments. In the deep learning approaches proposed in this work, we consider as input both these raw data, as well as we consider an approach that makes use of clustering experiments that generates a good parallelepiped surface for each atom, instead of the raw data. These strategies are implemented in terms deep feedforward networks and deep convolutional networks algorithms. For both architectures, the clustering-based strategy showed up promising results. However, models generated through deep feedforward showed up the best global results, being that those with the DBSCAN-Clustering-based approach highlight from the other ones.