Statistical vs. stochastic experimental design: an experimental comparison on the example of protein refolding

Optimization of experimental problems is a challenging task in both engineering and science. In principle, two different design of experiments (DOE) strategies exist: statistical and stochastic methods. Both aim to efficiently and precisely identify optimal solutions inside the problem-specific search space. Here, we evaluate and compare both strategies on the same experimental problem, the optimization of the refolding conditions of the lipase from Thermomyces lanuginosus with 26 variables under study. Protein refolding is one of the main bottlenecks in the process development for recombinant proteins. Despite intensive effort, the prediction of refolding from sequence information alone is still not applicable today. Instead, suitable refolding conditions are typically derived empirically in large screening experiments. Thus, protein refolding should constitute a good performance test for DOE strategies. We compared an iterative stochastic optimization applying a genetic algorithm and a standard statistical design consisting of a D-optimal screening step followed by an optimization via response surface methodology. Our results revealed that only the stochastic optimization was able to identify optimal refolding conditions (~1.400 U g(-1) refolded activity), which were 3.4-fold higher than the standard. Additionally, the stochastic optimization proved quite robust, as three independent optimizations performed similar. In contrast, the statistical DOE resulted in a suboptimal solution and failed to identify comparable activities. Interactions between process variables proved to be pivotal for this optimization. Hence, the linear screening model was not able to identify the most important process variables correctly. Thereby, this study highlighted the limits of the classic two-step statistical DOE.