The overall objective of this work has been to adopt new developments and techniques in the area of measurement, modelling and control of fermentation processes. Flow cytometry and software sensors are techniques which were considered ready for application and the focus was set on developing tools for research aiming at understanding the relationship between measured variables and process quality parameters. In this study fed-batch cultivations have been performed with two different strains of Escherichia coli (E.coli) K12 W3110 with and without a gene for the recombinant protein promegapoietin. Inclusion body formation was followed during the process with flow cytometric detection by labelling the inclusion bodies with first an antibody against the protein promegapoietin and then a second fluorescent anti-antibody. The approach to label inclusion bodies directly in disintegrated and diluted cell slurry could be adopted as a method to follow protein production during the process, although the labelling procedure with incubation times and washings was somewhat time-consuming (1.5 h). The labelling of inclusion bodies inside the cells to follow protein production was feasible to perform, although an unexplained decrease in the relative fluorescence intensity occurred late in process. However, it is difficult to translate this qualitative measurement into a quantitative one, since a quantitative protein analysis should give data proportional to the volume, while the labelling of the spheric inclusion bodies gives a signal corresponding to the area of the body, and calibration is not possible. The methods were shown to be useful for monitoring inclusion body formation, but it seems difficult to get quantitative information from the analysis. Population heterogeneity analysis was performed, by using flow cytometry, on a cell population, which lost 80-90% viability according to viable count analysis. It was possible to show that the apparent cell death was due to cells incapable of dividing on agar plates after induction. These cells continued to produce the induced recombinant protein. It was shown that almost all cells in the population (≈97%) contained PMP, and furthermore total protein analysis of the medium indicated that only about 1% of the population had lysed. This confirms that the "non-viable" cells according to viable count by cfu analysis produced product. The software sensors XNH3 and µNH3, which utilises base titration data to estimate biomass and specific growth rate was shown to correlate well with the off-line analyses during cultivation of E. coli W3110 using minimal medium. In rich medium the µNH3 sensor was shown to give a signal that may be used as a fingerprint of the process, at least from the time of induction. The software sensor KLaC* was shown to respond to foaming in culture that probably was caused by increased air bubble dispersion. The RO/S coefficient, which describes the oxygen to substrate consumption, was shown to give a distinct response to stress caused by lowered pH and addition of the inducing agent IPTG. The software sensor for biomass was applied to a highly automated 6-unit multi-bioreactor system intended for fast process development. In this way also specific rates of substrate and oxygen consumption became available without manual sampling. / QC 20100819

Orientadores: Ana Maria Frattini Fileti, Galba Maria de Campos Takaki / Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Quimica / Made available in DSpace on 2018-08-06T21:06:24Z (GMT). No. of bitstreams: 1 Albuquerque_ClarissaDaisydaCosta_D.pdf: 8871218 bytes, checksum: 55101c854d0f7293da9222f2518c7c15 (MD5) Previous issue date: 2006 / Resumo: Entre as técnicas convencionais usadas para desenvolvimento de software sensores, redes neurais artificiais têm mostrado ser um instrumento poderoso em modelagem e controle de bioprocessos complexos. O objetivo geral do presente trabalho é o desenvolvimento de software sensores baseados em redes neurais para estimação e previsão em tempo real da concentração de biomassa e da atividade de emulsificação no processo de produção de bioemulsificante por Cândida lipolytica. Bioemulsificantes normalmente apresentam vantagens como biodegradabilidade, baixa toxicidade e biocompatibilidade em relação a emulsifícantes sintetizados quimicamente. Adicionalmente, eles apresentam potencial para serem sintetizados a partir de substratos de baixo custo e são normalmente efetivos em condições extremas de pH, temperatura e salinidade. Consequentemente, bioemulsificantes têm sido aplicados com sucesso em áreas como biorremediaçào e recuperação de óleos. Contudo, os bioemulsificantes não são ainda largamente empregados por conta do seu alto custo de produção, resultante primeiramente da baixa produtividade dos microrganismos empregados e do alto custo de recuperação. Entretanto, neste trabalho foi mostrado que o desenvolvimento de softsensores neurais juntamente com otimização de componentes de meio de produção e ampliação de escala do processo podem contribuir para tomar a produção de bioemulsificante mais eficiente e econômica. O modelo quadrático obtido na otimização dos componentes do meio usando planejamento composto central com três fatores e metodologia de superfície de resposta mostrou significância estatística e capacidade preditiva. A máxima atividade de emulsificação para emulsões água-em-hexadecano obtida foi de 4,415 UAE e as concentrações ótimas de uréia, sulfato de amónio e fosfato monobásico de potássio foram respectivamente iguais a 0,544 % (w/v), 2,131 % (w/v) and 2,628 % (w/v). O rendimento do processo foi otimizado em 272%. A ampliação do processo da escala de frascos para a escala de fermentador de bancada foi realizada com sucesso e os efeitos e interações da temperatura e velocidade de agitação sobre a atividade de emulsificação do bioemulsificante produzido por Cândida lipolytica foram investigados. Os conjuntos de dados necessários para o treinamento, validação e teste dos softsensores foram obtidos de experimentos de produção de bioemulsificante realizados em biorreator de 5L, sob diferentes condições de temperatura e agitação. Os conjuntos de treinamento, validação e teste dos softsensores foram suavizados e expandidos usando interpolação com spline cúbica. Várias topologias de redes neurais com uma camada escondida foram testadas. As variáveis de entrada do processo incluíram pH, oxigênio dissolvido, densidade ótica e salinidade do liquido metabólico livre de células. O algoritmo de treinamento usado foi o algoritmo de retropropagação baseado em Levenberg-Marquardt em conjunção com regularização bayesiana. A raiz do erro quadrático médio (RMSE) e o coeficiente de determinação global (Rg2) foram usados entre outros índices para comparar o desempenho dos modelos. Os resultados mostram que softsensores neurais fornecem estimação e previsão on-line de concentração de biomassa e de atividade de emulsificação dentro de uma variação aceitável de 5% dos valores experimentais. Coeficientes de determinação global superiores a 0,90 indicam o excelente ajuste dos modelos de redes neurais com os valores experimentais testados, obtidos para concentração de biomassa e atividade de emulsificação / Abstract: Among conventional techniques used for development of 'software sensors', artificial neural networks have showed to be a powerful tool for modelling and control of complex bioprocess. The present work deals with the development of neural network based software sensors for real time estimation and prediction of biomass concentration and emulsification activity in a bioemulsifier production process by Candida lipolytics Bioemulsifiers commonly have the advantages of biodégradation, low toxicity, and biocompability over chemically synthesized emulsifiers. In addition, they can potentially be synthesized from cheap subtrates and are commonly effective at extremes of pH, temperature, and salinity. As a result, bioemulsifiers have found successful application in areas such as bioremediation and oil recovery. However, bioemulsifiers are not widely available because of their high production costs, which results primarily from low strain productivities and high recovery expenses. Therefore, in this work was showed that on-iine neural softsensor development jointly with media optimization and scale up of the process can make bioemulsifier production more efficient and more economical.The second order model obtained in the optimization of the medium components using three-factor central composite design and response surface methodology showed statistical significance and predictive ability. It was found that the maximum emulsification activity to water-in-hexadecane emulsion produced was 4,415 UEA and the optimum levels of urea, ammonium sulfate and potassium dihydrogen orthophosphate were, respectively, 0,544 % (w/v), 2,131 % (w/v) and 2,628 % (w/v). The emulsifier production process yield was optimized in 272 %. Successful scale-up from flasks to laboratory scale bioreactor was attained and the effects and interactions of the temperature and agitation rate on the emulsification activity of the bioemulsifier produced by Candida lipolytica were investigated. The data sets required to training, validation and test the neural software sensors were obtained from bioemulsifier production experiments carried out using com oil and sea water based media in a 5L bioreactor, under different temperature and agitation conditions. The training, validation and test sets were smoothed and expanded by interpolation using a piecewise smoothing cubic spline. Several neural network topologies with one hidden layer were tested. The input process variables included pH, dissolved oxygen, optic density and free cell metabolic liquid salinity. The training algorithm used was the Levenberg-Marquardt based backpropagation algorithm, in conjunction with Bayesian regularization. The root mean square error (RMSE) and the global determination coefficient (Rg2) among others index were used to compare model performances. The results showed that neural 'software sensors' supplied for biomass concentration and emulsification activity on-line estimation and prediction within an acceptable variation of 5% of the experimental values. Global coefficients of determination higher than 0.90 indicated excellent agreement of the neural network models with experimental test values, obtained for biomass concentration and emulsification activity / Doutorado / Sistemas de Processos Quimicos e Informatica / Doutor em Engenharia Química