SEAN RUANE*¹, EDUARDO LOPEZ-MONTERO², JASON FUNG², DARREN A. WHITAKER²
*Corresponding Author
1. Centre for Process Innovation, National Biologics Manufacturing Centre, United Kingdom
2. Perceptive Engineering – an Applied Materials Company, United Kingdom

Abstract

Integration of control models based on live data into development and manufacturing processes is a key development area in bioprocessing. This article describes data generation and statistical modelling for two bioprocessing unit operations, direct flow filtration and protein A capture chromatography. A Batch Wise Unfolding Partial Least Squares method was used to build models that could make predictions based on real-time data. For filtration a model was created that could predict filter blocking times based on current pressure trends with a 92% accuracy, and prescribe necessary interventions to operators. For chromatography a model was created which could predict process yields with a ± 5% accuracy to notify operators of out-of-spec performance.

INTRODUCTION
As the bioprocessing industry moves towards process intensification and continuous manufacturing, process robustness is of increasing importance, as process downtime results in greater product losses. Digital monitoring and controls based on real time process data provide a broadly implementable set of methods to increase process robustness, product safety and reduce product time to patients (1). Model-based predictive modelling and prescriptive maintenance is a key example, allowing detection and avoidance of process failures, increasing reliability and equipment utilisation and maintaining product quality. Predictive modelling can be used to predict process variables and outcomes in real time, such as process parameters or product quality attributes, allowing potential failure modes to be assessed. With model-based prescriptive maintenance real-time process data is used to diagnose failure modes and prescribe operator intervention. Such models provide a key step on the path towards advanced process controls driven by multivariate models, with the eventual goal being processes validated sufficiently to allow real-time product release (2).