Continuous glucose feedback control using Raman spectroscopy and deep learning models for biopharmaceutical processes.

Abstract

This study explores the implementation of continuous glucose control strategies in high-consumption, high-complexity cell culture processes using Raman spectroscopy and advanced deep learning models, including convolutional neural networks and variational autoencoder just-in-time learning. By leveraging deep learning-derived process monitoring, the study enhances glucose measurement accuracy and stability, enabling precise control across different glucose set points. This approach allows for a systematic evaluation of glycosylation effects and other critical quality attributes, addressing the impact of glucose variability on product consistency. Continuous glucose control is compared against traditional bolus feeding, demonstrating improved set-point maintenance, reduced high mannose (HM) levels, and enhanced overall titer productivity. To extend these benefits to manufacturing environments where Raman spectroscopy may not be feasible, a continuous glucose calculator (CGC) is developed as a scalable alternative. Experimental validation across multiple cell lines confirmed that both Raman-based and CGC-driven strategies minimized glucose fluctuations, reduced undesirable byproducts, and optimized process yields. These findings highlight the potential of continuous glucose control, combined with deep learning models, to improve bioprocess efficiency and product quality while addressing the challenges of dynamic, high-consumption bioreactor systems.