Near-infrared spectroscopy coupled with convolutional neural network as a checkpoint tool for cell culture bioprocess media characterization.

Abstract

As per the quality by design (QbD) paradigm, manufacturers are expected to identify critical raw materials that can contribute to variability in process performance and product quality. Further, manufacturers should be able to characterize and monitor the quality of these critical raw materials. Cell culture medium is universally accepted to be one such critical raw material for monoclonal antibody production. It is complex and comprises hundreds of components in varying proportions that are known to impact a multitude of critical quality attributes of a biotherapeutic product, particularly the post-translational modifications. In this study, a near-infrared (NIR) spectroscopy-based quantification method has been developed for media additives that are known to be potential glycan modulators. A one-dimensional convolution neural network (1D-CNN)-based chemometric model has been developed for estimating galactose and uridine concentrations in the various media formulations. Employing the advantage of data augmentation, the proposed 1D-CNN model delivers excellent prediction statistics (test R² > 0.9) for predicting both analytes in real time. Further, this model has been used in combination with DoE-based experimental design for prediction of glycosylation using concentrations of media additives as input. In summary, predicted glycosylation distributions were in accordance with actual distribution without significant differences (p > 0.9) in the investigated media formulation. The proposed method and tool can play a critical role in facilitating real-time characterization and control of mammalian cell culture raw materials.