In silico Metabolic Flux Data Flexibilization for Advanced Bioreactor Control Applications

摘要

The usual bioreactor control approaches are not reliable at low oxygen concentrations. So, the integration of machine learning along with systems biology may provide an alternative to overcome obstacles like the lack of sensibility of standard dissolved oxygen sensors. In this work, simulated metabolic data were used to obtain an Artificial Neural Network (ANN) that could be used as a simplified version of the metabolic model for online control purposes. Several growth conditions regarding oxygen limitation were run in silico using the Optflux software and the IND750 genetic scale model for yeasts. All the obtained in silico data was used to train and evaluate several structures of ANN. The best ANN architecture was later applied to experimented data for validation. An ANN with two hidden layers and 10 neurons each could successfully learn the respiratory quotient patterns for maximal ethanol production by Saccharomyces cerevisiae. This ANN structure also correctly predicted the experimental data used for validation. This surrogate model can be easily further applied in micro-aeration control strategies.