In-line prediction of viability and viable cell density through machine learning-based soft sensor modeling and an integrated systems approach: An industrially relevant PAT case study.

Abstract

The biopharmaceutical industry is shifting toward employing digital analytical tools for improved understanding of systems biology data and production of quality products. The implementation of these technologies can streamline the manufacturing process by enabling faster responses, reducing manual measurements, and building continuous and automated capabilities. This study discusses the use of soft sensor models for prediction of viability and viable cell density (VCD) in CHO cell culture processes by using in-line optical density and permittivity sensors. A significant innovation of this study is the development of a simplified empirical model and adoption of an integrated systems approach for in-line viability prediction. The initial evaluation of this viability model demonstrated promising accuracy with 96% of the residuals within a ±5% error limit and a Final Day mean absolute percentage error of ≤5% across various scales and process conditions. This model was integrated with a VCD prediction model utilizing Gaussian Process Regressor with Matern Kernel (nu = 0.5), selected from over a hundred advanced machine learning techniques. This VCD prediction model had an R² of 0.92 with 89% predictions within ±10% error and significantly outperformed the commonly used partial least squares regression models. The results validated the use of these models for real-time in-line prediction of viability and VCD and highlighted the potential to substantially reduce reliance on labor-intensive discrete offline measurements. The integration of these innovative technologies aligns with regulatory guidelines and establishes a foundation for further advancements in the biomanufacturing industry, promising improved process control, efficiency, and compliance with quality standards.