Inferência de variáveis do processo de produção de penicilina G acilase por Bacillus megaterium ATCC-14945.

• Autor(a) principal: Silva, Rosineide Gomes da
• Data de Publicação: 2003
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/3893
• topic: Engenharia bioquímica; Técnicas de inferência; Penicilina G Acilase; Modelagem matemática; Redes neurais; ENGENHARIAS::ENGENHARIA QUIMICA

The enzyme penicillin G acylase (E.C.3.5.1.11) is used in the production of 6-aminopenicillanic acid (6-APA) and 7-aminocephalosporinic acid (7-ACA), which are key intermediates for the production of β-lactam antibiotics. Its industrial importance was one of the motivations for this thesis. On-line measurement and monitoring of cultivations of Bacillus megaterium ATCC 14945 in a agitated and aerated bioreactor allowed the acquisition of variables such as mol fraction of CO2 and O2 in the exhaust gases, pH, dissolved oxygen. Batch and fed-batch runs, using either enzymatically hydrolyzed casein or a pool of free amino acids provided information concerning the extend of the cultivation, preservation of the microorganism and addition/exclusion of nutrients. Usual carbon sources as glucose, fructose and glycerol increased the cellular mass, but did not improve the productivity of PGA. The use of amino acids resulted in a 2.5-fold increase of productivity. Adding phenyl acetic acid at the beginning of the experiments did not inhibit cell growth. Three non-structured models of microbial growth were put forth, assuming one, two and three limiting substrates. However, these models were too simple for our purposes. Another approach was then followed, using neural networks (NN) as softsensors. Before implementing the inference algorithm, the blank noise of the instrumentation had to be reduced. A non-conventional filter was developed, combining a recursive NN, a moving average and a second recursive NN. The smoothed variables were the input for a second NN, for pattern recognition, which classified the run in one of the main growth phases: lag, exponential and stationary. The main purpose of this NN was to identify the exponential phase, which would be the domain of the next NN, a multilayer perceptron (MLP) for inference of the cellular mass. Several NN, with different topologies, were tested for this purpose. Finally, the product concentration (PGA activity in the medium) was estimated through a hybrid approach, using the growth rate inferred by the MLP NN, coupled to the cell-product yield, obtained from the fitting of the non-structured models. Another important information for this last algorithm was the knowledge that production of PGA was growth-associated, but with a 2hr-delay. The algorithm for inference was robust and accurate.