Machine learning modelling for the high-pressure homogenization-mediated disruption of recombinant E-coli

For French press mediated Escherichia coli cell disruption, the prediction capabilities of response surface method (RSM), multiple linear regression (MLR), multi-layer perceptron (MLP) and sequential minimal optimization (SMO) models are compared. The cell disruption process was analyzed as a function of three input variables, i.e. cellmass concentration (g/L), pressure (psi) and number of passes. The experimental data of the responses (nitrilase release, total protein release and extent of cell disruption) were employed to train above mentioned machine learning models. For nitrilase activity and extent of cell disruption, MLP model exhibited good generalization capability while for total protein release, SMO proved to be the best amongst others. The generalization capabilities of MLP and SMO models were almost close to that of RSM models as observed from error analysis of unseen data. Both nitrilase activity and protein release response were largely affected by the biomass concentration being disrupted while the extent of the cell disruption was heavily influenced by the operating pressure. In a single pass, at 12,000 psi, almost 22.3 U/mL nitrilase could be released for 300 g/L cellmass concentration after homogenization in a French-press.