Experimental design for fermentation media development: Statistical design or global random search?

The diversity of combinatorial interactions of medium components with the metabolism of cells as well as the large number of medium constituents necessary for cellular growth and production do not permit satisfactory detailed modelling. For this reason, experimental search procedures in simultaneous shaking flask experiments are used to optimise fermentation media. As an alternative to the methods of statistical experimental design employed in this field for many decades, the use of stochastic search procedures has been evaluated recently, since these require neither the unimodality of the response surface nor limitations in the number of medium components under consideration. Genetic algorithms were selected due to their basic capability for efficient exploration of large variable spaces. Using a genetic algorithm, it has been experimentally verified, with the aid of process examples, that process improvements can be achieved both for microbial and enzymatic conversions and for cell cultures despite the large number of medium components under simultaneous consideration (about 10 or more). In exploring a new variable space, process improvements of more than 100% were generally achieved. For initial reaction conditions previously 'optimised' via standard procedures it has been possible in most cases to achieve a further improvement of 20-40% of the target quantity. Although the genetic algorithm can be very efficient for exploration of large variable spaces, it is improbable that a 'global optimum' can be precisely identified because of the relatively small number of shaking flask experiments usually performed. As a consequence, a combination of highly directed random searches to explore the n-dimensional variable space with the genetic algorithm and subsequent application of classical statistical experimental design is recommended for media development.