Automatic Calculation of Locally Weighted Partial Least Squares (LW-PLS) will Improve Cell Culture Process Monitoring

 

Raman spectroscopy technology is a non-destructive, rapid, and robust method to measure multiple analytes simultaneously. That is why Raman Spectroscopy has become an essential Process Analytical Technology (PAT) tool in cell culture bioprocesses to monitor in-line and in real-time Critical Process Parameters (CPPs).

Partial Least Square (PLS) regression is the most common statistical method used to build chemometric models from Raman spectra to monitor CPPs in cell culture processes. To ensure models robustness, it is important for models to be calibrated with samples which are representative of the variability usually experienced by the process. Currently, in order to obtain an optimal coverage of a desired design space for a new cell culture process a simple Design-Of-Experiment (DOE) can be used to define the accurate measurements to be extracted from at least three batches.

A current goal in the field is to shorten the time dedicated to the generation of data needed to build models for a new bioprocess. An idea could consist in accumulating a wide range of data coming from batches achieved with a variety of cell lines and media, and in building generic models with them. However, recent experiments demonstrate that generic models provide poor accuracy. Then, how can we use generic datasets to easily build – ideally automatically – accurate models?

The use of Locally Weighted Partial Least Square (LW-PLS) models allows to perform automatically the selection of accurate samples in the database. The prediction accuracy of these models strongly depends on the definition of the relationship between a new sample and the samples stored in a database. To establish this relationship, similarity indexes have been used for the past 20 years. Similarity indexes have been usually defined on the basis of the Euclidean distance or the Mahalanobis distance. However, new other similarity indexes have emerged, based on angles between a query and samples in a database, on regression coefficients of PLS or on the covariance between input and output variables.

As model building is time consuming, the next step consists in on-line calculation of the LW-PLS used as a Just-In-Time (JIT) learning method. This effective tool of learning and prediction will enable to build more accurate models faster and will consequently improve the way of monitoring new cell culture processes.

 

References

Cleveland, S. Devlin, Locally weighted regression: an approach to regression analysis by local fitting. Journal of the American Statistical Association 83 (1988) 596–610.

Kim, M. Kano, H. Nakagawa, S. Hasebe, Estimation of active pharmaceutical ingredients content using locally weighted partial least squares and statistical wavelength selection, International Journal of Pharmaceutics 421 (2011) 269–274.

Hazama, M. Kano, Covariance-based locally weighted partial least squares for high-performance adaptive modeling, Chemometrics and Intelligent Laboratory Systems 146 (2015) 55–62.

Mei, Y. Ding, X. Chen, Y. Chen, Cai J Soft sensor modelling based on just-in-time learning and bagging-PLS for fermentation processes, Chemical Engineering Transactions, 70 (2018) 1435-1440T.

Webster, BC. Hadley, W. Hilliard, C. Jaques, C. Mason. Development of generic raman models for a GS-KOTM CHO platform process, Biotechnology progress, 34 (2018) 730-737.

Rowland-Jones, F. van den Berg, AJ. Racher, EB. Martin, C. Jaques. Comparison of spectroscopy technologies for improved monitoring of cell culture processes in miniature bioreactors, Biotechnology progress, 33(2017) 337-346.

 

European Commission logo

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 779218