In-line prediction of viability and viable cell density through machine learning-based soft sensor modeling and an...

Area of Science:

• Biopharmaceutical Manufacturing
• Process Analytical Technology (PAT)
• Systems Biology

Background:

• The biopharmaceutical industry is increasingly adopting digital tools for better systems biology data analysis and product quality.
• Implementing digital technologies streamlines manufacturing by enabling rapid responses, reducing manual work, and fostering automation.

Purpose of the Study:

• To develop and validate soft sensor models for predicting Chinese Hamster Ovary (CHO) cell viability and viable cell density (VCD) in real-time.
• To integrate an empirical model for in-line viability prediction with a machine learning model for VCD prediction.

Main Methods:

• Utilized in-line optical density and permittivity sensors to develop a simplified empirical model for viability prediction.
• Employed Gaussian Process Regressor with Matern Kernel (nu=0.5) for VCD prediction, selected from over 100 machine learning techniques.
• Integrated viability and VCD models within a systems approach for continuous monitoring.

Main Results:

• The viability model achieved high accuracy, with 96% of residuals within ±5% error and a Final Day mean absolute percentage error (MAPE) of ≤5%.
• The VCD prediction model demonstrated strong performance with an R² of 0.92 and 89% of predictions within ±10% error.
• Both models significantly outperformed traditional partial least squares regression methods.

Conclusions:

• Validated the use of integrated soft sensor models for real-time, in-line prediction of viability and VCD.
• Demonstrated potential to significantly decrease reliance on labor-intensive offline measurements.
• Established a foundation for advanced biomanufacturing, improving process control, efficiency, and regulatory compliance.