Modelling of bioprocess non-linear fluorescence data for at-line prediction of etanercept based on artificial neural...

In the last years, regulatory agencies in biopharmaceutical industry have promoted the design and implementation of Process Analytical Technology (PAT), which aims to develop rapid and high-throughput strategies for real-time monitoring of bioprocesses key variables, in order to improve their quality control lines. In this context, spectroscopic techniques for data generation in combination with chemometrics represent alternative analytical methods for on-line critical process variables prediction. In this work, a novel multivariate calibration strategy for the at-line prediction of etanercept, a recombinant protein produced in a mammalian cells-based perfusion process, is presented. For data generation, samples from etanercept processes were daily obtained, from which fluorescence excitation-emission matrices were generated in the spectral ranges of 225.0 and 495.0 nm and 250.0 and 599.5 nm for excitation and emission modes, respectively. These data were correlated with etanercept concentration in supernatant (measured by an off-line HPLC-based reference univariate technique) by implementing different chemometric strategies, in order to build predictive models. Partial least squares (PLS) regression evidenced a non-linear relation between signal and concentration when observing actual vs. predicted concentrations. Hence, a non-parametric approach was implemented, based on a multilayer perceptron artificial neural network (MLP). The MLP topology was optimized by means of the response surface methodology. The prediction performance of MLP model was superior to PLS, since the first is able to cope with non-linearity in calibration models, reaching percentage mean relative error in predictions of about 7.0% (against 12.6% for PLS). This strategy represents a fast and inexpensive approach for etanercept monitoring, which conforms the principles of PAT.