文摘
An artificial neural network (ANN) modeling scheme has been constructed for theidentification of both recombinant tissue-type plasminogen activator (r-tPA) proteinproduction and glycosylation from Chinese hamster ovary (CHO) cell culture, cultivatedin a stirred bioreactor. A series of hybrid feed-forward backpropagation neural networkswere constructed to function as a software sensor. This enabled predictions of viablecell density, r-tPA content, and r-tPA glycosylation. The sensor was based on an initialinput vector space consisting of simple metabolite concentrations, batch cultivationtime, and a description of shear stress applied to the culture. Metabolite concentrationsof the culture supernatant, included in the input vector space, were obtained from asingle isocratic HPLC measurement. The shear stress component of the input spaceenabled accurate culture state prediction over a wide range of agitation rates.Coefficient of determination (r2) values between ANN predicted and experimentalmeasurements of 0.945, 0.943, 0.956, and 0.990 were calculated to validate individualANN prediction accuracy for total ammonia, apparent viable cell density, total r-tPA,and Type II glycoform concentrations, respectively.