Pattern recognition of historical fermentation data for optimization of recombinant protein production

Using a fermentation system of Escherichia coli producing Green Fluorescent Protein (GFP), we have developed computational tools to utilize process data. This work focuses on learning patterns within historical data in order to optimize various outputs of the processes. The presented optimization schemes are based upon the generalized three-step method: (1) important input variables within the data base are identified, (2) the relationship between these critical inputs and a given process output is modeled, (3) the developed model is used to suggest novel input conditions that will optimize the process output. This generalized three-step optimization routine is designed to be robust towards complex nonlinear relationships and interactions between process inputs and outputs. First, the basic three-step optimization routine is implemented on a historical E. coli database generated in our lab. We show that this generalized three-step method increased the protein concentration by 55% compared to what had been previously observed. Time dependent fermentation models are then integrated into the generalized three-step method. This time dependent version of the three-step method is then implemented on a different historical E. coli database to predict optimal input conditions along with time dependent strategies. This time dependent approach was unable to increase the total protein yield. However, it did increase a simplified economic criterion by 15%. The successful implementation of this time dependent model was dependent upon Bayesian parameter estimation techniques. Novel Bayesian parameter estimation techniques were developed, and it was shown that a time-dependent fermentation model trained from Bayesian methods outperforms a model trained from using traditional weighted least squares parameter estimation. Last, experimental design criteria were developed that can be used in conjunction with the three-step method. These criteria can be utilized when there is presently insufficient data to optimize the system. However, enough information is available upon which to base future experiments upon that will eventually lead to an optimum. We show by using an E. coli database that these criteria can successfully identify regions of process input space that will lead to new optima.