Towards Enhancing Manufacturing Process Performance Through Multivariate Data Mining

Abstract

Several newly approved protein-based therapeutics in the past decade are manufactured in modern production plants with automated systems for process control and comprehensive data archival. The hundreds of process parameters and key output variables for several production batches in the vast historical databases provide a valuable resource to improve process understanding and robustness. Multivariate data analysis is a critical process analytical technology tool to unearth any hidden patterns within process trends and identify key parameters for enhancing process performance and product quality. Cell culture process data from more than hundred “trains” comprising production as well as inoculum bioreactors was investigated in this study. Each batch encompasses over 130 on-line and off-line temporal parameters. A maximum margin support vector algorithm was coupled with a kernel-based machine learning approach to develop multivariate predictive models for critical cell culture performance parameters. A differential weighting scheme was incorporated in the model to prioritize the process parameters with strong associations with process outcome and to identify key performance indicators at every stage of the production train. Model evaluations indicate that cell culture performance can be accurately predicted several days before harvest and downstream purification. Further, multiple parameters in the inoculum and early stages of production bioreactors were identified as precocious markers of the final process outcome. This process-data driven approach for knowledge discovery in manufacturing processes represents an important step towards implementing a real-time decision making scheme based on critical product and process traits.