Advancing bioprocess monitoring: data fusion and ANN-based prediction of arginine concentration in monoclonal antibody-producing CHO cell cultures.

In this study, we investigated the application of multivariate modelling approaches to extract quantitative information on arginine, a critical amino acid, from inline Raman spectroscopy during therapeutic monoclonal antibody-producing CHO cell cultivation. A central focus of the work was the implementation of a data fusion strategy, in which process-related information and dielectric spectroscopy-based viable cell density measurements were integrated with Raman spectral data to enhance predictive performance. This multimodal approach enabled more robust and reliable monitoring of arginine concentrations compared to single-sensor modelling. Partial Least Squares Regression (PLSR) and feedforward Artificial Neural Networks (ANN) were compared for their ability to develop calibration models from Raman spectra. While both methods demonstrated the potential for real-time arginine concentration monitoring, ANN models consistently showed better predictive performance (RMSEP of PLSR: 342.9 µM; ANN: 295.4 µM), particularly in terms of robustness in capturing concentration changes under different feeding strategies. The incorporation of fused process information into the ANN model further improved prediction accuracy, highlighting the advantages of ANN-based modelling combined with data fusion for advancing inline monitoring and supporting improved process control in biopharmaceutical production.