Coping with complexity: machine learning optimization of cell-free protein synthesis

Biological systems contain complex metabolic pathways with many nonlinearities and synergies that make them difficult to predict from first principles. Protein synthesis is a canonical example of such a pathway. Here we show how cell-free protein synthesis may be improved through a series of iterated high-throughput experiments guided by a machine-learning algorithm implementing a form of evolutionary design of experiments (Evo-DoE). The algorithm predicts fruitful experiments from statistical models of the previous experimental results, combined with stochastic exploration of the experimental space. The desired experimental response, or evolutionary fitness, was defined as the yield of the target product, and new experimental conditions were discovered to have ∼ 350% greater yield than the standard. An analysis of the best experimental conditions discovered indicates that there are two distinct classes of kinetics, thus showing how our evolutionary design of experiments is capable of significant innovation, as well as gradual improvement.