A Framework for Digital Technologies Application in Soft Sensing Monitoring to Recombinant Nanobody Production in <i>Escherichia coli</i> : From Model Explanation to the Study of Potential Scenario

The production of recombinant nanobody-based (protein) products in <i>Escherichia coli</i> exhibits promising results for its application in serotherapy, for example, the production of antivenoms against scorpion stings. Nevertheless, the monitoring, control, and optimization of this bioprocess present limitations due to the lack of measurements of critical variables. This work studies the application of digital technologies, such as machine learning algorithms, as soft sensors to estimate hard-to-measure variables. Six machine learning algorithms were employed to train and test the soft sensors in estimating protein and biomass concentrations: Classification and Regression Tree, Random Forest, k-Nearest Neighbors, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Multilayer Perceptron Neural Network (MLP). Online variables measured by sensors and two simulated variables derived from a dynamic hybrid model for recombinant nanobody production in <i>E. coli</i> were used to train algorithms. SHapley Additive exPlanations values were used as a post hoc tool to improve the interpretability of the soft sensor's estimations. SVM and MLP exhibited the best performance for the protein soft sensor, with training and test results of <i>R</i> ² > 0.81, MSE < 0.046, RMSE < 0.215, and MAE < 0.107. XgBoost, SVM, and MLP showed the best performance for biomass, with training and test results of <i>R</i> ² > 0.98, MSE < 2.370, RMSE < 1.539, and MAE < 1.154. Soft sensors were evaluated as a tool to study sensor faults by implementing committees of parsimonious soft sensors to graphically compare the estimations of protein and biomass concentrations. The results demonstrated the feasibility of applying soft sensors to real-time monitoring of key variables.