(293d) Application of a Raman Spectroscopy Analyzer and Process Data Analytics Tools to Enable in-Line Monitoring of Perfusion Mammalian Cell Cultures

Guided by the USA FDA recommendations on adaptation of Quality by Design (QbD) guidelines and activities of the International Committee for Harmonization (ICH), manufacturing of biologics has undertaken its biggest transformation in the last two decades. Process Analytical Technologies (PAT) applied to biomanufacturing have represented an enabler to the QbD paradigm to generate adaptive processes that can reduce product quality deviation through the implementation of sensor technology for process monitoring and control. This goal has also been benefited from the rapidly increasing digital transformation exploiting advances in data analytics pipelines, automation, increased computational power and enhanced management of data repositories.

Mammalian cell cultures are widely used as the workhorse platform in the production of biological products including antibodies, other therapeutic proteins, and growth factors. Deviations of critical process parameters (CPP) in upstream unit operations can directly impact physicochemical and biological heterogeneities in the biological expression system and associated critical quality attributes (CQA) in therapeutic product. Therefore, monitoring and control strategies are typically deployed using a wide range of available measurements (off-line, at-line, in-line and on-line) to ensure minimal variation within the bioreactor environment and its impact on cellular metabolism and product profile. Nevertheless, established on-line measurements for relevant process parameters are limited to sensors for pH, temperature, dissolved gases (pO2, pCO2), system pressure, capacitance, and off gas stream analysis. Other variables important to these platforms, such as extracellular concentrations of media nutrients, by-products and formed product along with product quality attributes, require the use of at-line and off-line measurements. Depending on the required analytical technique, this process can result in long delays and laborious methods for obtaining process descriptors and manual process corrective actions. Hence, there is significant value in tools supporting real-time monitoring for key cell culture parameters and product quality attributes that enable and facilitate bioprocess optimization and subsequent control.

Vibrational spectroscopic techniques, including mid-infrared (MIR), near-infrared (NIR) and Raman spectroscopy, represent molecular-specific analytical technologies with increasing use in the biopharmaceutical industry. Particularly for bioprocessing, Raman spectroscopy has become attractive for PAT applications given its inherent properties such as non-contact, non-destructive, high molecular specificity, and weak water bands for good quality analysis in aqueous solutions. Given the increasing interest for robust process design, optimization and control in emerging intensified and continuous upstream platforms, Raman spectroscopy provides a great potential for real-time and in-situ measurement of relevant cell culture process parameters and product quality attributes. Herein, we present a case study on the implementation of Raman spectroscopy and associated data analytics for process monitoring implementation in bench scale CHO cell perfusion bioreactors.

A set of perfusion culture runs was conducted using a monoclonal antibody secreting CHOZN^® GS^-/- ZFN-modified CHO cell line (MilliporeSigma, USA) in bench scale (3L) stirred tank bioreactors coupled to a novel TFF based cell retention device (Cellicon^TM Perfusion Solution, MilliporeSigma, USA). The perfusion operation with commercially available chemically defined media (EXCELL^® Advanced HD perfusion media, MilliporeSigma, USA) was controlled using a fixed cell specific perfusion rate (CSPR) with increasing cell density. Cultures were operated during at least eight cultivation days, reaching viable cell densities up to 170x10⁶ cells mL^-1 and cell viabilities above 97%. During cultivation, Raman spectra measurements were acquired (45min sampling interval with 30 spectra average) using an in-line single-channel ProCellics^TM Raman Analyzer (MilliporeSigma, France) with its optical probe immersed in bioreactor and equipped with a 785 nm excitation stabilized laser (300 cm^-1 to 4000cm^-1 bandwidth acquisition).

For data pre-processing, an investigation via in-silico full factorial design of experiments (in silico DoE) was conducted to understand the effect of parameter selection and implementation order of signal filtering techniques on high cell density culture datasets, including standard normal variate (SNV) transformation, Savitzky-Golay (SG) derivatives and spectral truncation. Pre-processed spectral data were modelled via Principal Component Analysis (PCA) to identify outliers and Partial Least Squares (PLS) to assess the different pre-processing combination on spectral regions.

Data processing yielded combinations of parameters that resulted in better performing preprocessing. These optimal combinations were used to generate and evaluate univariate and multivariate PLS models based on non-linear iterative partial least squares (NIPALS) method (Bio4C^TM PAT Raman Software, MilliporeSigma, France; SIMCA^® Software, Sartorius Stedim Biotech, Sweden; MATLAB, The Mathworks Inc USA). Models showed high explained (R²Y>0.90) and predicted (Q²>0.90) variance for most parameters with reduced prediction accuracy for ammonia ions due to difficulty for decorrelation. Using a validation data subset, not used previously during calibration stage, different univariate and multivariate PLS models were configured in terms of latent variables and evaluated using root mean squared error in predictions (RMSEP) with generation of predicted and observed values curves for different target features. Given the constraints of PLS as a linear-based multivariate tool, other non-linear supervised learning regression techniques (Support Vector Machines and Random Forests) were utilized to benchmark the PLS model performance during training and validation stages (MATLAB, The Mathworks Inc USA; R, R Foundation for Statistical Computing, Austria). Feature importance was evaluated for each technique to assess differences in prediction performance. Overall, the prediction accuracy of these Raman based multivariate models were estimated below ±10% for all cases of target parameters, being comparable with the accuracy of the off-line analyzers and hence representing the potential for implementation of in-line monitoring.

Overall, this investigation introduced an experimental and multivariate modelling approach for the adoption of Raman spectroscopy for in-line monitoring in perfusion mammalian cell cultures. The different presented data handling techniques at the pre-processing and processing stages demonstrated their potential to build robust predictive multivariate models for relevant cell culture process features based on Raman spectra acquisition. Consequently, the implementation of combined advanced sensor technology and data analytics tools can enable process monitoring, and subsequent control, to improve process yields and meet product quality requirements in next generation intensified upstream platforms.