458703 Multivariate Data Analysis (MVDA) for Cell Culture Process Development

Thursday, November 17, 2016: 2:36 PM
Continental 5 (Hilton San Francisco Union Square)
John Bowers, BioProcess Development, Merck & Co., Kenilworth, NJ, John Higgins, Merck and Company, Inc., West Point, PA, Maria Khoiuzam, TECHNOLOGY CENTER, Merck, Kenilworth, NJ, Louis Obando, Manufacturing Division, Process Analytical Technologies (PAT), Merck and Company, Inc., West Point, PA and Balrina Gupta, Upstream Process Development, Merck, Kenilworth, NJ

Multivariate Data Analysis (MVDA) for Cell Culture Process Development

John Bowers, Maria Khouzam, Louis Obando, John Higgings, and Balrina Gupta.


Analyzing mammalian cell culture results has traditionally been done after batch completion by creating and reviewing univariate plots of each variable separately. This approach results in a large amount of data that is not easily interpreted. MVDA Multivariate Data Analysis (MVDA) resolves these because it analyzes parameters and their covariance of large data sets to summarize the performance of the process. Real-time MVDA allows for online process monitoring, and provides immediate understanding.

To apply MVDA during cell culture development, several hurdles were overcome. The number and variety of bioreactors generated our first challenge. Experiments were performed in systems ranging from an automated set of twenty four 250 mL mini-bioreactors to a manufacturing skid with two 500 L stainless steel bioreactors. Since batch context was not available in a consistent format across the bioreactor controllers, special software was developed to enter it into PI. Furthermore each reactor had its own configuration, so a universal configuration structure was develop and the actual configuration of each bioreactor entered into it.

In addition, continuous process data were collected into the data archiving system while off-line results were collected into a separate electronic notebook (ELN) system, making it harder to use combined results to holistically analyze new batches. To combine offline measurements with continuous bioreactor data, SIMCA-Online was configured to extract the daily offline measurements from E-Notebook BioAssay® (PerkinElmer) database. Batch information and the continuous data were taken from a PI data historian. A “grab and hold” of the last entered value was sufficiently representative of daily offline measurements to account for the sampling rate difference between the daily and continuous measurements in the model. Models based on these combined modes significantly better described culture behavior, and were much better predictors of batch quality measures, such as titer or charge variants.

Furthermore, modeling approaches were developed to account for scale changes and intentional parameter adjustments.  Most parameters used were intrinsic, such cell and metabolite concentrations, pH, DO, temperature etc. To account for scale effects in useful extrinsic parameters, such as gas flows or feed amounts, a volume based scaling was introduced. With these, we were able to use results from automated 250 mL mini-bioreactor system Ambr® 250 (Sartorius Stedim Biotech) to build models for 3-L and larger scale bioreactors.



In order to account for intentional changes in parameters, such as pH or temperature settings, variables were tracked as deviations from setpoints.  The deviation from setpoint allowed for diagnostics of equipment performance, without the multivariate statistics being dominated by an intentional setpoint change. We applied this approach to process ranging studies to determine the impact of changing a parameter on overall process performance live, giving immediate feedback on what a parameter change impacted. When needed, the setpoints were then included in the models to quantify the magnitude of the change to performance or quality attributes.

We successfully applied MVDA using SIMCA-13 and SIMCA-Online® (Umetrics) to several development activities, such as assessing the impact of the parameter changes on the process performance, confirming similarity between master and working cell banks, and demonstrating process comparability across scales. Applying the techniques described here, as much information as possible was obtained from development activities, even in the context of continuing process changes. Six to eight initial similar batches can be used to create models for future work. These models, in turn, can be used to improve process understanding to guide future work. The new results can then be included in models to further improve understanding and model performance.


Extended Abstract: File Not Uploaded