Article
BIOINFORMATICS
DOI: 10.2478/v10133-010-0092-5
PRO-OPTIMIZER: A NOVEL WEB-ENABLED OPTIMIZATION ENGINE FOR MICROBIAL FERMENTATIONS E.B. Gueguim Kana1, J.K. Oloke2, A. Lateef2 and A.F. Donfack Kana3 1 University of KwaZulu-Natal, School of Biochemistry, Genetics and Microbiology, Pietermaritzburg, Scottsville, South Africa 2 Ladoke Akintola University of Technology, Department of Pure and Applied Biology, Ogbomoso, Nigeria 3 Ahamadu Bello University, Department of Mathematics, Zaria, Nigeria Correspondence to: E.B. Gueguim Kana E-mail:
[email protected]
ABSTRACT
Traditional strategies designed to determine the best combination of physico-chemical parameters to optimally drive a fermentation process suffer from large computational burden, or are unable to account for the interaction effects of the various process parameters. In this paper, we report a web-enabled engine for fermentation optimization using Genetic Algorithm. The setpoint values of the process parameters are concatenated into strings, called chromosomes and represent potential solution states. A population of such strings is formed. These chromosomes are evaluated for their fitness using an objective function or through experimentation. The best-fitted chromosomes are selected and undergo mutations and crossovers to produce the next generation. This procedure is repeated until the fitness value becomes constant and the best chromosome emerges. The system considers the interactive effects of all the process parameters. The effectiveness of this engine has been evaluated on the feeding strategy for the production of Saccharomyces cerevisiae DS2155, optimization of yogurt acidification process using Lactobacillus bulgaricus and Streptococcus thermophilus. Data revealed a feeding strategy with a biomass yield improvement of 0.53g/l on S. cerevisiae, and an acidification slope of 0.06117 compared to an initial of 0.0342, reducing the yogurt acidification time from six to two hours. This system will reduce the cost for fermentation process development. Biotechnol. & Biotechnol. Eq. 2010, 24(4), 2137-2141 Keywords: fermentation, optimization, genetic algorithm, artificial neural network, distributed laboratory
Introduction
From the industrial point of view, process optimization is of critical importance because it determines the overall process economics. This may be defined as the control of a process at its optimal state and to reach its maximum productivity with the lowest possible cost, while quality is maintained (2). Traditional strategies for the optimization of fermentation processes suffer at least one major drawback. Component swapping, which consists of swapping one component for another one fails to consider the component concentrations and interaction effects (6). The biological mimicry or mass balance concept strategy which mimics the cell composition requires knowledge of detailed elemental composition of the cell; this composition differs from strains to strains and their growing environments. One Variable At a Time (OVAT), historically the most popular due to its ease of application and its ability to show individual component effect on the process has a flaw of ignoring the interactions among all process components. Fermentation process optimization based on mathematical models, described by a set of differential equations derived from mass balances (13), are costly (11). They lack robustness and accuracy due to the physiological complexity of microorganisms (5) and the dynamic optimization problems of such complex systems are difficult to solve. The conventional analytical methods, such Biotechnol. & Biotechnol. Eq. 24/2010/4
as Green’s theorem and the maximum (or minimum) principle of Pontryagin, are unable to provide a complete solution due to singular control problems (9). Meanwhile conventional numerical methods, such as dynamic programming (DP), suffer from a large computational burden and may lead to suboptimal solutions (7). Artificial intelligence tools such as Artificial Neural Network (ANN), Genetic Algorithm (GA) and Fuzzy Logic provide a new approach to address this class of problems. Complex optimization problems can be solved without knowing the impacts of each parameter in detail (12). Genetic algorithm (GA) is a computer optimization search strategy that mimics the biological evolution of best-fitted species (8). GAs search for optimal solution to a problem by mimicking the biological evolution of the best fitted species using genetic operations such as mutations, crossovers and selections. With GA, an optimization problem is treated through a cycle of four stages: creating of population of solutions, selection of the best solutions, evaluation of those solutions, selection of the best solutions and breeding using the parent population and genetic manipulation to create a new population of solutions. The cycle continues until a suitable result is achieved (1) (Fig. 1). For example, given a search of numerical values of 100 parameters used to define an optimal temperature profile for a nonlinear fermentation process, one can concatenate these values into a long string, called chromosomes or solution states by GA specialists. A population of 50 of such strings can be formed and called a generation. These chromosomes are evaluated 2137
for their fitness on the process using an objective function or through experimentation. The best-fitted chromosomes are selected and undergo mutations and crossovers to generate a new population. This protocol is repeated until the fitness value becomes constant and the best profile or chromosome emerges. An example of comparison of GA with dynamic programming (DP) for optimization has been given by Nguang et al. (10) on the determination of the optimal feed rate profile of a fedbatch culture of monoclonal antibodies (MAb). The result showed that a final production of MAb by using GA was about 24% higher than that produced using the DP. This paper reports the implementation of a web based software for optimization of chemical, biochemical and physical parameters of fermentation processes using Genetic Algorithm. In its optimization process, the interactions between all the process parameters are observed. The effectiveness of this engine has been evaluated on a number of fermentation processes, among which are the development of an optimum feeding strategy for the production of Saccharomyces cerevisiae (3) DS2155 and optimization of yogurt acidification process using Lactobacillus bulgaricus and Streptococcus thermophilus (4).
Fig. 1. General Flowchart of Genetic algorithm
Materials and Methods Genetic Algorithm structure for optimization search This module was structured on Darwin’s idea of evolution by the survival of best-fitted biological individuals, in which the individuals are substituted with solution states. A solution state 2138
is the best combination of various substrates concentrations or physical and biochemical parameters to drive the fermentation process. For example, S1=(35oC, 5 g/l, 0.01g/l, 7) and S2=(30oC, 10g/l, 0.01g/l, 3) could be two solution states for the setpoint values of temperature, glucose concentration, ammonia concentration and pH, respectively for a given fermentation process. In this example, S1 and S2 can be called chromosomes and the setpoint values of temperature, glucose concentration, ammonia concentration and pH are the genes. To search for an optimal solution state, the algorithm proceeds through the following steps: S1. A set of solution states from an initial search domain are randomly selected to form the generation 0 (G0). From generation 0, subsequent generations iteratively evolve after application of GA operations such as: S2. Mutations: Few setpoints on a given solution state are randomly substituted with some random values within the search range. S3. Crossover between solution states: Two randomly selected solution states undergo an exchange of a sequence of setpoints at a random site, S4. Performance evaluation: This consist in ranking all potential optimal solution state based on an objective evaluation or field experimentation, S5. Selection of the best solution states and iteration from S2. At each genration the performance of the best state increases, and also does the average performance of the generation. This is done iteratively until an optimal state emerges or a stopping criterion set by the user is met. This evolutionary algorithm was coded using the PHP script programming language and the database type was developed using Mysql. Software user interface The corresponding user interface for the programmed algorithm is made up two major panels, the process setup and the process data sheet panels. The process setup panel (Fig. 2) has two windows; the first is used to configure the GA parameters such as the generation size, crossover rate and mutation rate. The second window is used to input the optimization parameters such as the parameter’s name, their search range and the search steps. The search range specifies the minimum and the maximum values within which a potential optimal value can be expected. The step is the difference between two consecutive values. It reflects the search accuracy. For example, in the search of an optimum pH value within the range 3-8, one can use a step of 0.5, if a pH tolerance of 0.5 is acceptable for the process. The second panel is the experiment data sheet (Fig. 3). An experiment datasheet corresponds to a set of solution states to be evaluated either in the laboratory or using an objective function, and an input column for the evaluated performance. Each experiment data sheet corresponds to a generation of individual in Darwin context of evolution. Beside the process Biotechnol. & Biotechnol. Eq. 24/2010/4
Fig. 2. Process set up panel of the pro-optimizer software
experiment datasheet panel, there is a window for optimization performance monitoring. Detailed guidelines for users are presented within the user guide. Furthermore, the effectiveness of the Optimization search engine was assessed in the optimization of glucose feeding strategy for Saccharomyces cerevisiae DS2155 production and optimization of the acidification process in yoghurt production using Lactobacillus bulgaricus and Streptococcus thermophilus. For S. cerevisiae optimization, the feeding profiles were the solution states. The biomass productivity of the feeding profile was the performance index. The process setup has been reported elsewhere (3). An Artificial Neural Network model was developed based on experimental input/ output data. This model was used as objective evaluator to rank the solution state. On the Genetic Algorithm configuration panel, a crossover rate of 0.6, a mutation rate of 0.001 and a population size of 50 were used. With these parameters, the genetic algorithm iteratively searches through different generations using genetic operations, and ANN to predict the performance of various feeding profiles in each generation. The search process was stopped at the 6th generation when an optimal feeding profile emerged. With regard to the optimization of yoghurt production, the solution states were made of the temperature profiles varying every hour within a process time of 5 hours. The performance index was the acidification slope of the temperature profile. Using the optimization module, the crossover probability was set at 0.5, the mutation rate at 0.08 and population size was maintained at 30. GA sub-module randomly selected a set of temperature profiles from an initial set to form the generation 0 (G0). Subsequent generations iteratively evolved after Biotechnol. & Biotechnol. Eq. 24/2010/4
application of GA operations such as mutations, crossover on the profiles, ranking using their performance value and selection. This was done iteratively until an optimal profile emerged or the stopping criterion was met. Details on the fermentation process setup and the optimization search strategy have been reported (4).
Results and Discussion
The Genetic Algorithm used to implement this optimization engine is a multidimensional random search algorithm which ensures global convergence when given sufficient time. The process setup panel (Fig. 2) can take up to 20 parameters, thus allowing for complex optimization search in the fermentation process in which the interactive effects of these parameters on the process behaviour is considered. This outweighs a good number of current optimization tools. For example using the statistical factorial Design Of Experiment, 1024 experimental runs would be required if 5 parameters are considered at 4 levels. The web enabled feature of this engine allows the experiment data sheet (Fig. 3) for the same process to be assessed by different researchers at the same time. This brings the possibility of a process optimization in distributed laboratories, thus different researchers may optimize the same process from different geographically located laboratories, submitting, viewing data and observing the process behaviour in real-time. With S. cerevisiae, at each generation, the average performance as well as the performance of the best individual feeding profile improved. The evolved optimal feeding profile showed a predicted biomass productivity of 2.33g/l, an improvement of 0.53. The performance trend is 2139
Fig. 3. Experiment data sheet of the pro-optimizer software
shown in Fig. 4. This study revealed that a multiphase feeding strategy with values of specific growth rate oscillating near the critical growth rate value can be an alternative superior feeding strategy. For the optimization of the acidification process in yoghurt production using Lactobacillus bulgaricus and Streptococcus thermophilus the stopping criterion was met after 11 generations (Fig. 5). The optimal trajectory showed an acidification slope of 0.06117 compared to an initial of 0.0342 and a setpoint sequence of 43, 38, 44, 43, and 39oC. This emphasized the earlier experimental suggestion that process initialization at high temperature and termination at low temperature resulted in a faster acidification. Findings revealed that fast acidification could be achieved by initializing the process at 43oC and terminating it 39oc.
Fig. 5. Optimization performance trend for acidification of yogurt
Conclusions
Although the use of Genetic Algorithm in the optimization of fermentation process is not a completely new trick in the toolbox of bioprocess engineering, a web-enabled software system for fermentation process optimization using genetic algorithm, allowing individual process optimization to be done in distributed laboratories is being reported here for the first time. In all optimization cases evaluated with this system, an external Artificial Neural Network module has been used to develop a model which served as evaluator, thus substituting the objective function. These models were trained using experimental data and have proven to be able to capture the nonlinear behaviour underlying the fermentation process contained in the training data. Fig. 4. Optimization performance trend for S. cerevisiae
2140
Biotechnol. & Biotechnol. eq. 24/2010/4
This system eases the determination of optimal control setpoint values of physico-chemical parameters of industrial fermentations without an in-depth knowledge of the process’s dynamics. Pro-optimizer engine shortens the fermentation process development time.
REFERENCES
1. Baishan F., Hongwen C., Xiaolan X., Ning W., Zongding H. (2003) Process Biochemistry, 38, 979-998. 2. Chen L., Nguang S.K., Chen D.X., Li X.M. (2004) Biochemical Engineering Journal, 22, 51-61. 3. Gueguim E.B., Oloke J.K., Lateef A. (2007a) African Journal of Biotechnology, 6(9), 1122-1127. 4. Gueguim E.B., Oloke J.K., Lateef A., Zebaze M.G. (2007b) Journal of Industrial Microbiology and Biotechnology, 34(7), 491-496. 5. Hussain M.A. (1999) Artificial Intelligence in Engineering, 13, 55-68.
Biotechnol. & Biotechnol. Eq. 24/2010/4
6. Kenedy M. and Krouse D. (1999) Journal of Industrial Microbiology & Biotechnology, 23, 456-475. 7. Katare S. and Venkatasubramanian V. (2001) Eng. Appl. Artif. Intell., 14,715-726. 8. Michalewicz Z. (1996) Genetic Algorithm Data Structure Evolution programs, Springer-Verlag, Berlin. 9. Na J.G., Chang Y.K., Chung B.H., Lim H.C. (2002) Bioprocess Biosyst. Eng., 24, 299-308. 10. Nguang S., Chen L., Chen X. (2001) ISA Trans., 40, 381389. 11. Rani K.Y. and Rao V.R. (1999) Bioprocess Eng., 21, 7788. 12. Ronen M., Shabtai Y., Guterman H. (2002) J. Biotechnol., 97, 253-263. 13. Schugerl K. (2001) J. Biotechnol., 85, 149-173.
2141