Bioprocess Monitoring in the Biopharmaceutical Industry Using Filtering and Online Cost-Sensitive Supervised Learning

Abstract

This thesis explores the enhancement of bioprocess monitoring in biopharmaceutical production using advanced filtering techniques and cost-sensitive learning algorithms. The first study addresses the challenge of optimizing recombinant protein production in Escherichia coli (E. coli ) fermentations, utilizing Kalman, particle, bilateral, and extended Kalman filters to improve titer estimation accuracy. Using fermentation data from Cytovance Biologics, the proposed techniques effectively impute missing titer values and are validated on a secondary dataset from Chinese Hamster Ovary (CHO) cell lines, demonstrating their robustness and accuracy in bioprocess monitoring. In the second study, cost-sensitive learningalgorithms—specifically, variants of Passive-Aggressive (PA) and Cost-Sensitive Online Gradient Descent (CSOGD)—are applied to control chart pattern recognition (CCPR) for process control. These algorithms outperform their standard non-cost-sensitive counterparts across various abnormal control chart patterns, including uptrend, downtrend, upshift, downshift, cyclic, and stratification. The inclusion of class-specific penalties and surrogate hinge loss functions enhances classification performance, particularly in smaller window lengths (w ≤ 25). Additionally, cost-sensitive algorithms demonstrate stable performance across dynamic imbalanced datasets, though they incur higher computational costs due to increased complexity. Together, these studies provide a comprehensive framework for optimizing both the biological aspects of protein production and the operational stability of manufacturing processes, significantly advancing bioprocess monitoring in the biopharmaceutical industry.