Desenvolvimento de algoritmos evolucionistas para aprimorar a metodologia PSP de novo do programa Rosetta

• Ano de defesa: 2021
• Autor(a) principal: Werdt, Paulo Roberto Teixeira
• Orientador(a): Dardenne, Laurent Emmanuel
• Banca de defesa: Dardenne, Laurent Emmanuel, Ocaña Gautherot, Kary Ann del Carmen, Barbosa, Hélio José Corrêa, Pascutti, Pedro Geraldo
• Tipo de documento: Tese
• 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: Biologia molecular - Processamento eletrônico de dados; Algorítmos genéticos; Bioinformática; Proteínas
• Área do conhecimento CNPq: CNPQ::CIENCIAS BIOLOGICAS::GENETICA
• Link de acesso: https://tede.lncc.br/handle/tede/343

One of the main challenges of the computational molecular biology has been the development of methods capable of predicting the native structure of a protein from its sequence of amino acids. The Critical Assessment of Protein Structure Prediction has been documenting the progress and the state of the art of PSP. Nowadays PSP techniques are sub divided into two main groups: Template based, when proteins solved experimentally are used as starting points for the generation of new models; and Free Modeling (FM), where the generation of new models are obtained strictly from physicochemical principles (ab initio) or when some experimental information is used (de novo). Among the main problems that a PSP FM method has to face are: (i) the deceptiveness associated with the evaluation functions due to the theoretical simplifications introduced to make the model computationally viable; (ii) the size, complexity and multimodality of the search space, that result from the possible conformations associated with a given protein sequence. In many cases the native structure does not even belong to the global minimum basin of the investigated surface. The results of the evaluations of the CASP events indicate the Rosetta program as one of the most successful methods of PSP de novo. Its algorithm uses fragments of experimentally determined structures to guide the search in structural space. The insertion of these fragments is performed by a Monte Carlo algorithm, which assembles the fragments into the structure using Simulated Annealing, evaluated by mixed score functions, based both on molecular mechanics and on knowledge from databases of known structures. Its optimization strategy uses a large number of independent executions of the algorithm, where each trajectory seeks to quickly converge to a minimum of the energy surface, to obtain a large set of predictions, from which a subset of the most promising models is chosen. Literature results indicate that the Rosetta energy functions are the strong point of its methodology, being widely used in other successful PSP FM methods, yet its conformational search strategy can be improved. Bibliographic research on related works shows that, in fact, the Rosetta algorithm cannot efficiently explore the fragment's multimodal space, which reinforces the idea of using a different method to improve its conformational search. The objective of this work is to improve the conformational search of Rosetta through the development of evolutionary metaheuristics, in order to achieve a better mapping of the search space and obtain gains, both in the quality of the predicted structures as for the reduction of the high computational cost that Rosetta requires to get good quality models. In this approach, a populational algorithm with parental replacement by phenotypic crowding is used, where the offspring of candidate solutions replace the elements in the parental population by a criterion of structural similarity. Variations of the algorithms have also been developed with different population regimes and selection mechanisms: tournament based and by the structural clustering of individuals of the population, in addition to a crossover operator that considers secondary structure predictions. The implementation of these techniques to the Rosetta de novo methodology was gradually tested in experiments with a diverse group of protein targets. The results obtained with the most successful resulting algorithms have been contrasted with corresponding results from the original Rosetta, and also with two other similar approaches found in literature. They have demonstrated to reach a positive balance between the exploration and the exploitation of the energetic hypersurface, generated by the different scores used during the assembly of the fragments. Two of the algorithms proposed in this work achieved significant improvements, not only in the diversity and structural quality of the predicted final models, but also in the reduction of the computational cost necessary to obtain good quality structures.