Construção de Funções Empíricas Utilizando Rede Neural para Determinação de Constantes de Afinidade Receptor-Ligante

• Ano de defesa: 2008
• Autor(a) principal: Rêgo, Thais Gaudencio do
• Orientador(a): Dardenne, Laurent Emmanuel
• Banca de defesa: Raupp, Fernanda Maria Pereira , Sant'anna, Carlos Maurício 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: Serviço de Análise e Apoio a Formação de Recursos Humanos
• País: BR
• Palavras-chave em Português: Biomoléculas - Estrutura - Simulação por computador; proteínas - Estrutura - Simulação por computador; Atracamento molecular; Desenvolvimento de fármacos; Energia livre
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://tede.lncc.br/handle/tede/101

The understanding of receptor-ligand molecular recognition is one of the central aspects in structure-based design and discovery of new drugs. The key methodology is the docking of small molecules in active sites of proteins. There are two aims in any program of molecular docking: the search for the best ligand-protein conformation and the calculation of the free energy of this association, or its affinity constant. The test set used in this work was composed by 50 protein- ligand complexes, with experimentally measured Ki or Kd values for the construction of an empirical function specific to the DOCKTHOR program, using as input variables: energy of electrostatic interaction and Lennard-Jones, contact area of ligand-receptor on the surface accessible to the solvent, the presence of hydrogen bridges, and the number of the ligand rotatable bonds that were frozen in the process of docking. These variables were used for the construction of two types of free energy scoring functions. The importance of each variable used as input data for the construction of those functions was rated by means of multiple regression. A neural network was x also used to try to build the best model for the calculation of the affinity constant. The DOCKTHOR program currently has a prediction power leading to r = 0.4245, which indicates the importance of improving its scoring. The function built with the multiple regression methodology used 5 input variables and had linear, quadratic, and cross-product terms leading to r = 0.7542. Using the methodology of group cross-validation (VCG), it was concluded that the best architecture for the neural network consists of 9 neurons in the hidden layer, as it has the smallest error of generalization and greater consistency in errors. In the tests with this neural network architecture built using the same 50 protein-ligand complexes in training and test, 66% of the complexes had a difference smaller than 1.0 in the observed values. The generalization error (obtained by VCG) of a neural network that uses 9 neurons in the hidden layer was about ten times lower than that obtained by using a polynomial function. This is an indication of the superiority of the neural network methodology with respect to the multivariate regression methodology, specially for an empirical function developed for a broad range of receptor-ligand complexes.