High throughput optimization of medium composition for Escherichia coli protein expression using deep learning and Bayesian optimization

Abstract

To improve synthetic media for protein expression in Escherichia coli, a strategy using deep neural networks (DNN) and Bayesian optimization was performed in this study. To obtain training data for a deep learning algorithm, E.coli harvesting a plasmid pRSET/emGFP, which introduces the green fluorescence protein (GFP), was cultivated in 81 media designed using a Latin square in deepwell-scale cultivation. The media were composed of 31 components with three levels. The resultant GFP fluorescence intensities were evaluated using a fluorescence spectrometer, and the intensities were in the range 2.69-7.99× 103. A deep neural network model was used to estimate the GFP fluorescence intensities from the culture media compositions, and accuracy was evaluated using cross-validation with 15% test data. Bayesian optimization using the best DNN model was used to calculate 20 representative compositions optimized for GFP expression. According to the validating cultivation, the simulated GFP expression levels included large errors between the estimated and experimental data. The DNN model was retrained using data from the validating cultivation, and secondary estimations were performed. The secondary estimations fit the corresponding experimental data well, and the best GFP fluorescence intensity was 1.4-fold larger than the best of the initial test composition.