DEVELOPMENT OF AN ANALYTICAL PLATFORM FOR PAT MONITORING OF AN ETANERCEPT BIOPROCESS THROUGH AN INTEGRATED CHEMOMETRIC MODELLING APPROACH

CHIAPPINI, FABRICIO A.; ALCARAZ, MIRTA RAQUEL; AZCARATE, SILVANA M.; FORNO, ÁNGELA G.; GOICOECHEA, HÉCTOR CASIMIRO; TEGLIA, CARLA MARIELA

Lisboa

Encuentro; ESACT Meeting European Society for Animal Cell Technology; 2022

Background and novelty: Process Analytical Technology (PAT) constitutes a methodological framework promoted by Food and Drug Administration (FDA) that encourages the pharma industry to implement high-performance analytical techniques, increasing the efficiency in process design, monitoring and control. In PAT development, chemometrics plays a fundamental role in biochemical data modelling.In this work, a mammalian cell-based bioprocess for etanercept production was studied, focusing on two main issues. Firstly, at the end of the fermentation step, the cell culture experiments an unpredictable drop in cell viability. Hence, the assessment of this variable through cell counting may not provide information sufficiently in advance to take corrective actions. Secondly, as the product concentration is monitored through an off-line univariate chromatographic method (HPLC), a considerable delay usually occurs between sampling and results. The purpose of this investigation was to develop a PAT platform for the at-line monitoring of cell viability and etanercept concentration.Experimental approach: Both aspects were tackled through the modelling of multiway fluorescence data, which were generated from daily fermentation samples and modelled through a variety of chemometric techniques. Principal component analysis (PCA) was used for exploratory study and pattern recognition, partial least squares-discriminant analysis (PLS-DA), for multivariate classification, and multilayer perceptron-artificial neural network (MLP), as a non-parametric method for multivariate calibration.Results and discussion: PCA of spectral data revealed a differential and consistent pattern regarding the evolution of cell viability. In all lots, the last sample considered of high viability according to cell counting appeared closer to samples with lower viability. Therefore, spectral information was able to anticipatedly alert the moment of viability drop, since it can be associated with non-observable metabolic changes. These findings motivated the development of a multivariate classification model, aiming to obtain a predictive tool for the prospective inference of cell viability in future lots. Non-error rates of 100% in both training and prediction datasets were obtained. On the other hand, fluorescence and HPLC data were used to train an MLP model, which proved to be a suitable regression strategy due to data non-linearity. This algorithm was rationally optimized through response surface methodology (DoE-RSM). The optimum model yielded a percentage relative error of prediction of about 7%. Finally, it should be emphasized that both qualitative and quantitative prediction strategies arise from the generation of a unique type of analytical data, which conforms PAT requirements.