Utilização de rede neural para predição de proteínas de bactérias secretadas por vias não clássicas

• Ano de defesa: 2020
• Autor(a) principal: Oliveira, Luiz Gustavo de Sousa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Redes neurais; Proteínas bacterianas; Secreção por via não clássica; Neural Networks; Non-classical Secretory Pathway; Bacterial Proteins; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Computação
• Link de acesso: https://repositorio.ufu.br/handle/123456789/31262; http://doi.org/10.14393/ufu.di.2021.34

We present a methodology of predicting proteins from the non-classic bacterial secretion from artificial neural networks to contribute to vaccines’ elaboration and diseases’ diagnosis. We compiled a list of bacterial proteins from prokaryotic organisms secreted by the currently known non-classic pathways for training. These proteins were cataloged, allowing the creation of a set of training and validation for neural network training. We carried out bibliographic research to identify probable descriptors and signaling characteristics of this type of secretion by non-classic pathway. We developed a supervised neural network using the WEKA software, training it to determine the best group of features for prediction. We evaluated our proposed method submitting proteins not used in the training group and comparing the developed predictor against two other predictors studied in related literature, PeNGaRoo (ZHANG et al., 2020) and SecretomeP 2.0 (BENDTSEN et al., 2005). We considered our results satisfactory, as they presented a balanced neural network accuracy of 93% in the classification performance. We outperformed SecretomeP 2.0 for all validation scenarios. In the majority, our results were similar to PeNGaRoo, but for some case studies, we outperformed it. Therefore, we demonstrated the possibility of obtaining a compelling classifier by using our selected set of descriptors.