Process Analytical Technology within Biotechnology

MSc Chemistry

Analytical Sciences

Literature Thesis

Process Analytical Technology within Biotechnology

by

S. Sangers

10442138

July 2016

12 ECTS

April – July 2016

Supervisor/Examiner:

Examiner:

Dr. J.A.Westerhuis

Dr. W.Th. Kok

S. Sangers

1

Summary

This literature study is focused on how product quality of biopharmaceuticals can be achieved using Process Analytical Technology (PAT). In 2004 the Food and Drug Administration (FDA) released their framework for implementing PAT and Quality by Design (QbD) for pharmaceutical manufacturing processes. Since then, biopharmaceutical manufacturers have made significant advances to implement QbD and PAT on their complex production processes. And by doing so, they try to move slowly out of the traditionally Quality by Testing (QbT) paradigm. These QbD and PAT setups have led to increased understanding of the production process, and could potentially lead to building the product quality into the manufacturing process. This product quality is defined by its critical quality attributes (CQAs) which are ideally directly monitored by PAT. However, direct analysis of these attributes is often an analytical challenge, and therefore correlated critical process parameters (CPP’s) are analyzed instead.

This thesis will start off with the typical steps to be taken to implement PAT and QbD for a biopharmaceutical production process. First, a short introduction regarding a typical biopharmaceutical production process is given, including with a selected overview of available PAT sensors. Secondly, multivariate data analysis on explorative experiments is covered, these analyses are required to increase process understanding and to establish the CQA’s and the correlating CPP’s. Subsequently multivariate model construction of a production process is reviewed. Such a model can be used to monitor the process in real time with the appropriate PAT sensors for the CPP’s or even measure the CQA’s directly. These models should be well trained and validated in order to make them capable to predict the production trajectory. And can then be used to identify if a process will go out of spec, which could lead to insufficient product quality. This capability can then be used for real time intervention of the process in such a way to bring the process back to the correct trajectory. Finally, the current status of PAT and QbD implementation with the field of biopharmaceutical manufacturing will be discussed. In this discussion, it was found that there are numerous papers presenting results of relative small scale case studies.

S. Sangers

2

initial steps to be taken towards QbD and PAT implementation. Such as the explorative analysis, determining the CQAs and CPPs and even monitoring the manufacturing processes using multivariate models. Sadly however, actual regulatory filings of a QbD based process are limited and there are no known publications of a production process utilizing real time release testing. Hence, challenges still remain such as real time releasing product batches and applying remedial process steps after a fault has been detected. In addition, it was also found that the industry is struggling with the regulatory adjustments required for implementation of a QbD production process. Furthermore, the industry is reluctant to implement QbD practices due to high costs and initial high complexity of introducing the required technology. Therefore, QbT remains the common path for biopharmaceutical manufacturing and filing these processes and products to the regulatory agencies.

S. Sangers

3

1 Table of Contents

Summary ... 1

2 Abbreviations ... 4

3 Introduction ... 5

4 Biopharmaceutical process and product understanding ... 7

4.1 The production process ... 7

4.2 PAT sensors ... 8

4.3 Product Quality ... 9

5 Quality by Design ... 10

5.1 Explorative analysis, process knowledge ... 11

5.2 Monitoring the process ... 13

5.3 Prediction of process outputs ... 15

5.4 Control of future process steps... 16

6 Applications ... 17

7 Discussion ... 21

S. Sangers

4

2

Abbreviations

2D-HPLC Two Dimensional High Performance Liquid Chromatography

CHO Chinese Hamster Ovary

CPP Critical Process Parameter

CQA Critical Quality Attribute

DO Dissolved Oxygen

DoE Design of Experiments

FDA Food and Drug Administration

HPLC High Performance Liquid Chromatography iVCC integrated Viable Cell Count

ivPCV integrated Viable Packed Cell Volume

mAB Monoclonal antibody

MLPCA Maximum Likelihood Principal Component Analysis MVDA Multivariate Data Analysis

NIR Near Infrared

NOC Normal Operating Conditions

OD Optical Density

PAT Process Analytical Technology

PCA Principal Component Analysis

PCV Packed Cell Volume

PLS Partial Least Squares regression

QbD Quality by Design

QbT Quality by Testing

SCL Superior Control Limit

SPR Surface Plasmon Resonance

SVR Support Vector Regression

S. Sangers

5

3

Introduction

Traditionally product control in biotechnology is accomplished by Quality by Testing (QbT) where raw materials, samples of in-process materials and end products from a product batch are tested to evaluate if the quality of these materials and final drug product meet the FDA approved specifications or other predefined standards.1 These specifications are often derived from small scale development experiments, and they may not scale properly with an increase scale of production processes, which might lead to out of spec end drug product. As the manufacturers have the production process approved by the FDA, they cannot simply adjust process parameters without filing supplements with the FDA.1 Thus, biopharmaceutical products are released under the QbT paradigm, in which each batch will be subject to multiple release tests which a product batch needs to pass. When one of these tests fails, it leads to product discard. By developing and operating a production process under QbD, a systematic approach that emphasizes on product and process understanding, real time release testing during the production process might be possible. This could eliminate the need for release testing on the final drug product. Additionally, as the entire production process is better understood, being monitored and controlled by utilizing Process analytical technology (PAT). The process can even be intervened to correct a faulty process trajectory, saving a product batch from a discard.1,2

Utilizing PAT has great potential as it can help to better understand and control the production processes and is essential for QbD implementation. The FDA acknowledged this when they released their PAT – framework,3 in which they unbendingly recommend the use of Quality by Design (QbD) and PAT when manufacturing products. By applying these tools the FDA visions that Quality is built into products. The FDA sees PAT as “a system for designing, analyzing, and controlling manufacturing

through timely measurements (i.e., during processing) of critical quality and performance attributes of raw and in-process materials and processes, with the goal of ensuring final product quality”.3

Successfully implementing QbD and PAT could lead to a tremendous amount of benefits which include but are not limited to: improved process understanding, continuous process improvement, improved information for regulatory submissions, less production failures and possibly real time release of batches.4–6 Moreover, QbD enables increased flexibility in the manufacturing processes in a post-approval setting, thus potentially allow more efficient, faster and likely cheaper production processes.6 PAT is all about measuring a variety of parameters during the different phases of the production

S. Sangers

6

temperature,7,8 or two dimensional data such as near infrared spectra (NIR).9–11 At first PAT was applied in the production of small molecule pharmaceutical ingredients,12,13 nowadays PAT technology is gaining increased interest in the production of biopharmaceuticals,14 as their market share is expanding rapidy.15 The biotech production arena has its own unique set of challenges, as the production quality of the therapeutic proteins is easily altered by minor changes within the environmental parameters of the production process such as, pH, nutrient concentrations, waste products variants, etc.16 Moreover, direct measurements inside a bioreactor is challenging due to the complex nature of its contents.14

This literature study is focused on how product quality of biopharmaceuticals can be achieved by implementing PAT and QbD, as the FDA suggests within their framework document.3 It will start off with a brief introduction regarding a typical production process of a biopharmaceutical. Next, an overview of a selection of typical PAT sensors (and their capabilities) that can be utilized during the fermentation step of biopharmaceutical production process will be given. Next data interpretation will be discussed, primarily focusing on the role of MVDA. Subsequently the current implementation of PAT, QbD and MVDA in a few recent applications will be reviewed, in order to evaluate the feasibility of QbD and PAT for biopharmaceutical production processes. Finally, the current status of QbD and PAT implementation within the field of biopharmaceutical products is discussed, since the release of the QbD and PAT guidance framework by the FDA.

S. Sangers

7

4

Biopharmaceutical process and product understanding

This chapter will be used to give a brief description of a typical biopharmaceutical production process. Additionally, it will provide a selected overview of sensors that can be used in a PAT based production process. Next, measuring biopharmaceutical product quality will be highlighted, this quality is determined by CQA’s. As analysis of these attributes is typically an analytical challenge, correlation with CPP’s will be investigated and how these procedures can be incorporated into a PAT and QbD based production process.

4.1 The production process

Biopharmaceuticals are produced by recombinant techniques. In which, cell lines are genetically engineered to produce the target biopharmaceutical.17 This production process is typically divided into two sections, upstream and downstream processing, see Figure 1. In the upstream section the fermentation takes place, starting in relative small bioreactors to finally scaling towards 20.000 liter bioreactors.17 After the harvest step of these cell cultures, the downstream part the process takes place. Here, the biopharmaceutical is purified and formulated as drug substance.18 This formulated drug substance is subjected to off-line batch release testing14, before it is released for clinical or commercial use.

Although, PAT and QbD can be applied on both the upstream and the downstream parts of the production process, this thesis will mostly be focused at the upstream section of the production process. As this step has a large impact on the product yield and quality. Currently, control of the process and the harvest time is defined by rigid standardization.14 If operation could take place on a QbD basis this would potentially allow for more process control with regard to product yield and quality.

S. Sangers

8

Figure 1 Schematic manufacturing process of monoclonal antibodies from cell culture, divided in an upstream (USP) and a downstream (DSP) section. In the USP part the fermentation is scaled up to the final large scale bioreactor where the cell culture produces the desired antibody. In the downscale scale process the antibody is harvested and further purified into drug substance.19

4.2 PAT sensors

To build a QbD upstream production process the bioreactors should equipped with suitable sensors to measure various parameters.20 Typical parameters of interest during the fermentation process are: nutrients concentrations, pH, temperature, dissolved oxygen concentration, etc. Measurement of these parameters can be challenging as the contents of bioreactors are often a complex mixture of cells, nutrients, waste products, etc. However, continuous advances in sensor technologies have led to feasibility of real time measurements of many bioprocess parameters.21,22 PAT sensors can also be utilized at the downstream section of a biopharmaceutical production process. This phase of the production process is focused on the yield of the product from the consecutive purification steps.23,24 For example, some researchers investigated monitoring column resin materials, by doing so they were be able to predict resin decay of a protein A column, which allowed for timely replacement and leave

S. Sangers

9

be equipped at-line (next to the process analyzing samples), on/in-line (within a continuous (side-) stream of sample) or in-situ (inside a vessel itself, i.e. a probe). An overview of a selection of available sensors during the upstream phase is given in Table 1, where the measurement principle, the typical application and the PAT state are listed. In this table the PAT state implies whether the sensor is applied off-line, at-line, on-line or in-situ.20

Table 1: Overview of PAT sensors which can be equipped at the bioreactor to measure a variety of parameters during the upstream phase in order to monitor the production process of a therapeutic protein.

Sensor Measurement principle Application PAT state Reference Surface plasmon

resonance Refractive index change Product concentration and affinity At-line Jacquemarte et al. (2008)27

Near infrared (NIR) Transmission at various

wavelengths

Detect media components At-line Hakemeyer et al. (2012)9

High performance

liquid chromatography Physicochemical properties Amino acid and glucose concentration On-line Larson et al. (2002)28

High performance

liquid chromatography Physicochemical properties Product concentration and structure (CQA) On-line Rathore et al. (2010)29

RF impedance based Capacitance of membrane bound biovolume

Determine biomass On-line Carvell et al. (2005)30

In situ microscopy Optical imaging of

cell population Characterize cell population In-situ Joeris et al. (2002)

31

Flow cytometry Light scattering and

fluorescence Characterize cell population At-line Kacmar et al. (2005)32

In situ 2D fluorometry Fluorophore

emissions Detect media components and metabolic end products

In-situ Teixeira et al. (2008)33

4.3 Product Quality

Although PAT sensors can measure a variety of parameters of the production process, product quality is often not covered by these parameters. As product quality is defined by the critical quality attributes (CQAs) such as: glycosylation, charge variants, 3D structure and the level of aggregation.19,34 These product attributes are of great importance for product’s safety and/or efficiency.4 Some of these CQA’s are challenging to measure by using at-line/on-line techniques. For example the glycosylation

S. Sangers

10

a turnaround time of 5 hours,35 while Burnina et al. managed to get the turnaround time down to 90 minutes,36 with both of the authors applying 96 well plates to enable high throughput analysis. Another example of an analytical challenging CQA are the charge variants of a biopharmaceutical, typically this parameter is analyzed off-line using ion exchange chromatography.37 In a recent study reported by Amand et al.,38 a two dimensional high performance liquid chromatography (2D-HPLC) method was developed for a mAb which was able to operate at-line. The assay consists of a protein A purification step of 100 µL sample from the bioreactor as the first chromatographic dimension, subsequently the purified monoclonal antibody (mAb) is analyzed by separating the different charge variants by weak cation exchange chromatography.

Nevertheless, CQAs remain an analytical challenge to measure by utilizing PAT sensors. However, CQA’s are often correlated to certain critical process parameters (CPP’s).14,34 For example a study by Le et al., showed that the lactate concentration is correlated towards the antibody titer in their mAb production process.11 Another study by Konno et al.39 showed that the osmolality of the cell culture was correlated to the fucose content of the glycosylation of mAbs. And Grainger et al.40 reported correlation of N-glycosylation with manganese, galactose and uridine. In such cases, measuring these CPP’s is more feasible than the CQA’s in an at-line/on-line setting. Therefore, increased process understanding can be vital to identify the CPP’s which are correlated towards the analytical challenging CQA’s.34,41

5

Quality by Design

Moving from the traditional QbT based processes towards QbD based processes, increased process and product knowledge is essential. Therefore the product and process are analyzed for many parameters, either with off-line reference methods or with PAT sensors. These measurements of all the different parameters result in big data sets, which require proper interpretation. Using traditional linear univariate regression techniques can result in misleading interpretations and conclusions.41–43 Therefore, Multivariate data analysis (MVDA) should be applied when handling these large data sets and their multidimensional nature.14,41,44–46 To begin, it should be applied to explore existing data sets (preferably from a design of experiments (DoE)) of the production process, which can help in identifying the variables with the largest impact on the process. Next to that, MVDA can model the process and

S. Sangers

11

be utilized for online monitoring of the process. Finally, the monitoring model can be used for fault detection, deviation from the predicted trajectory, and remedial procedures could be applied to correct the process back to the predicted trajectory.34,41,43,45

5.1 Explorative analysis, process knowledge

To achieve QbD based product processes, increased process and product knowledge are essential. Therefore, data should be acquired by a systematic approach of experiments using a varying set of parameters. Preferably, the experiments are setup via a Design of Experiments (DoE) to enable full benefit of the following multivariate data analysis.46 With this structured approach, MVDA can be applied to increase the knowledge of the production process and the most impacting process parameters can be revealed,34,47 as illustrated in Figure 2. Here a set of experimental productions are performed and these are analyzed by omics-techniques such as, metabolomics, fluxomics, proteomics, transcriptomics and genomics. As the bio-production processes are cell based, in-depth understanding regarding the cell dynamics is essential in order to control product quality and quantity.48 When all the combined data is analyzed with MVDA, it gives insights in the complex interactions of the process media with the biological systems of the cells.34,41 Thus, making the production process within the cell more predictable and better controllable.48,49 This increased process knowledge aids to identify the CQA’s and to which process parameters these CQA’s are correlated. Subsequently this data can be used to establish a knowledge space where the process is relatively well understood. Within this knowledge space, the design space is located (Figure 2). Within this design-space, the process would typically be operating (also known as Normal Operating Conditions (NOC)), producing within this space should lead to good quality products.

S. Sangers

12

Figure 2: Improving process understanding of therapeutic protein production. High-throughput omics technologies and Design of Experiments (DoE) knowledge space the design space of all critical process parameters (CPPs) required to meet the critical quality attributes (CQAs) of a therapeutic protein. The gathered process knowledge combined with advances in analytical tools will allow to gradually improve the monitoring and control of the manufacturing process, to maintain all CPPs within the predefined ranges.34

Although a DoE setup of data is preferred, early process development regularly offers numerous experiments with variable parameter settings. It is interesting to analyze those data by MVDA. Though, these data sets are unstructured, incomplete and truncated. For example Mercier et al.43 reported successful analysis of data sets containing production data of batches during different time lengths. In that study it was decided to exclude some of the data to obtain a workable dataset, see for illustration Figure 3. Subsequent MVDA on the remaining data was capable to reveal interesting features of the production process by interpretation of the PCA and PLS score plots.

This explorative analysis of the production process leads to a Design space/NOC. Subsequently, monitoring and controlling the production process by PAT can be implemented. This will be further elaborated in the upcoming paragraphs.

S. Sangers

13

Figure 3 Schematic representation of data multiway unfolding. In option 1, only batches longer than 11 days were

considered, the shortest batches are excluded. In option 2, all batches are truncated at 7 days of culture. For both options, data excluded are shaded.43

5.2 Monitoring the process

To effectively obtain production process data for MVDA, implementation of on-line or at-line measurements of CPP’s or directly CQA’s are a necessity. Later on, this gives the possibility to detect faults in real time and correct the production process timely in order to ensure product quality.41 For instance, Clavaud et al.50 monitored large scale cultivation of mammalian Chinese hamster ovary (CHO) cells producing a monoclonal antibody. Using NIR measurements in combination with MVDA they were able to predict 7 parameters (protein content, iVCC, Glucose concentration, PCV, VCD, ivPCV and osmolality) simultaneously with a high degree of accuracy. These predictions were subsequently utilized for in-line monitoring of the product titer and the integrated viable cell count (iVCC).50 This approach showed that using NIR can effectively monitor several parameters in real time and thus reduce the use of conventional analytical tools for these parameters.

In another case study reported by Kirdar et al.51, the feasibility of MVDA was investigated for supporting the manufacturing of a biopharmaceutical. Using the data of 6 input parameters (pCO2, pO2,

S. Sangers

14

viability, and osmolality) from 14 small-scale batches (2L), they constructed a PLS model. Subsequently, this model was used to generate control charts, Figure 4 shows one of the control charts generated by Kirdar. It shows the batch trajectory of their 14 small-scale batches and a ± 3× standard deviation control bands. The production trajectories of the individual batches show that the production process within this selection of batches is well controlled and no faults are observed.

Figure 4: Batch control chart for the small scale batches, including 3x standard deviation control bands (red). If a batch trajectory is violating the control bands, it can be reasonably expected that the batch performance is abnormal. 51

MVDA can also be used for fault detection, for example Gunther et.al.52 showed that PCA can be used to discriminate between industrial fed-batch cell culture from normal operation conditions and the batches from deviating operating conditions. Data from 20 NOC batches and 3 abnormal batches was available. They constructed a PCA model using a total of 19 NOC batches and 14 process parameters. After cross-validation a PCA model with three principal components was generated. Subsequently this model was used to determine 99% confidence limits for Q (Sum of Squared Residuals) and T2 (Hotelling's T2). Their model constructed on NOC data could successfully detect the abnormal batches based on these 99% confidence limits.

In one more study by Nucci et al. 53 a PCA model was constructed based on MVDA data of the production process of penicillin G acylase. Online monitoring of the following parameters was

S. Sangers

15

gas, inlet air flow rate and stirring speed. As output parameters they determined the enzyme activity and the cell concentration. Within their control charts, a Superior Control Limit (SCL) was calculated and added to the PCA control charts. This enabled fault detection capabilities as shown in Figure 5, the left panel shows the process trajectory of a flawless run, while the right panel shows there is an abnormal run. The faults are easily detected as the process trajectory crosses the SCL multiple times.

Figure 5: On-line monitoring charts of the PCA model by Nucci et al., including the SCL. Violating the SCL means a fault is detected. The left panel shows the process trajectory of a run without faults. The right panel shows a process trajectory with a fault.53

5.3 Prediction of process outputs

The next challenge is to construct a MVDA model which has predictive capabilities regarding the CQA’s or other key parameters, for instance the final product titer11,50 _{and lactate concentration.}11 _When

a sufficient large data set is available it could be used to construct and train a PLS model. For instance, Le et al.11 constructed a model to predict the final product titer and the final lactate concentration based on a dataset of 234 runs. While Streefland et al.10 managed to make predictions on the expected product quality (based micro array gene expression data) at the end of a cultivation process. In Chapter 6 more details regarding these two applications are given.

S. Sangers

16

model (as outlined in §5.2) as tool to predict the batch trajectory of large-scale batches (2,000 L) based on a model which consist of data from small-scale batches.51 This prediction deems successful and thus implies that using a limited amount of data from early small-scale development batches can be used to predict the process on a much larger scale.

In the case study by Clavaud et al.,50 multiple cultivation parameters are predicted based on a PLS model constructed by using 10 independent batches. Seven input parameters were monitored by an in-line NIR sensor, subsequently the constructed PLS model was able to predict the process trajectory.

5.4 Control of future process steps

Finally, the constructed model should be connected to the PAT setup, in such a way that the production process can be controlled. A few case studies show models capable to control and steer the production process. For instance, Le et al.11 reported that they could predict the final product titer based on lactate concentration earlier in the process. Using this knowledge, they suggest that intervention and remedial of potential low productivity runs is possible, resulting in a substantial increase in productivity. They found that lactate consumption was the prominent factor in determining the final product titer within this production process. In addition, this result of their research suggests opportunities to steer the metabolism towards high productivity. However, they also found that potential interventions to steer the process should be executed at the inoculum or the early stage of the production scale reactors. Although they did show the potential of steering the production for higher production yields, it was not demonstrated. More details regarding this paper are given in Chapter 6.

In another paper by Hakemeyer et al.,9 a PCA model was constructed based on NIR measurements from Chinese Hamster Ovary (CHO) production processes at different scales. This model was used to predict critical process parameters and batch trajectory. This predicted data was then plotted onto control charts and compared with the actual data from the production process. By utilizing this feature, they were able to closely monitor and potentially correct the process in case some fault is detected. For instance, when a batch trajectory crosses the confidence limits, some intervention might be required in order to get the process back to the expected trajectory, for instance by adjusting the feeding strategy

S. Sangers

17

adjusted feed strategies based on their PAT model.

In a paper by Navratil et al.54 a PCA model was constructed based on the data from laboratory scale fermentation processes. This model was calibrated using data from reference methods such as: Optical density (OD) to measure the biomass, Glucose and acetate concentrations by on-line HPLC and the product titer (recombinant cholera toxin B) was measured using surface plasmon resonance (SPR). The data was generated by NIR spectroscopy and an electronic nose. Control charts were made based on the model which could be used to monitor the process trajectory and deviations from normal behavior could easily be identified. Additionally, the model could accurately predict the concentration of the formed cholera toxin B. Comparing these predicted growth rates to the actual grow rates determined by PAT sensors, gave the capability to steer the fermentation by adjusting the glucose feed rate.

6

Applications

Within this chapter a two papers will be elaborated more extensively. These are reports of case studies which have applied all the previous described steps and try to show the potential benefits of QbD and PAT implemented at the production process of a biopharmaceutical. These potential benefits may include, increased product yield or quality and better process understanding.

Streefland et al.10 reported full PAT implementation at a bacterial vaccine cultivation process. Using a DoE matrix of 12 experiments they determined the design space of this cultivation process. By combining NIR data (to indicate specific attributes of interest and as a “fingerprint” that gives qualitative information regarding the process) with conventional cultivation process data (i.e. pH, temperature etc.) into a single database, they constructed the model describing the multivariate process. The product quality was determined by microarray assay where they calculated a product quality score based on the genes from virulence core regulon.55 As input for the MVDA they used the following:

 The setup parameters for the DoE (T, pH, Dissolved Oxygen (DO), preculture density, inoculation density and reactor).

 The time varying process variables (pH, DO, T, flow of air, O2, and N2, stirrer speed).

S. Sangers

18

 At-line data of Optical density (OD), glutamate and lactate concentration.

 The product quality score from the microarray assay.

Their MVDA consists of 4 steps, first the design space was determined using multiple linear regressions. Subsequently a PLS model was constructed from the obtained at-line NIR data, which was capable to predict process variables. This PLS model showed that the dissolved oxygen concentration could be reasonably predicted. Additionally, this PLS model showed good predictive capabilities with regard to the growth curve, which was plotted in real time, minimizing the need for excessive sampling. Finally, the NIR data was used to construct a PCA model using five principal components, explaining 99.9% of the variation. This percentage of the variation seems exceptionally large, as other report lower coverage of the variation explained. The model by Mercier et al.43 explained 89% of the variance with 5 principal components, while Teixeira et al.33 reported models explaining 85.4 – 99.8% of the variation using up to 14 principal components. Nevertheless, these five components were used as a PCA fingerprint that contains abstract characteristics of the product and process quality. Still it is assumed that the fingerprint is similar between batches and can thus be used for non-biased batch-to-batch comparisons. Finally, by applying an approach described earlier by Wold et al.56 a model was constructed which could be used for monitoring the cultivation process. All their produced batches were concluded to be of good quality, as none of the batches were outside 95% confidence interval of the score plot of the first two components. However, they acknowledge that the amount of data is insufficient to allow real-time release of the product, therefore additional data from new batches should be added to the model, especially from data near the edges of the design space.10

While Streefland et al.10 showed that MVDA analysis of PAT data of a small dataset can provide a model to monitor the production process, they were unable to real time release their product. The review of the next paper by Le et al.11 will show how a model can be constructed using a large data set, hundreds of runs containing many variables. They showed that this model can be used to predict and control the final titer of the product at the end of the process using correlation with the lactate concentration.

In this paper, Le et al.11 constructed a PLS and a SVR model to predict the product and lactate titer from a data set of 243 runs containing 134 process parameters. Within this dataset there are runs at four scales of bioreactors (80 L, 400 L, 2,000 L and 12,000 L). These different scales are the result of the

S. Sangers

19

these datasets were also segregated into several stages: up to 70 h, 120 h, 170 h, 220 h, and 260h. This was finally organized into eight individual and seven cumulative datasets (i.e. data of 80L + 400L, and 80 L + 400L + 2,000 L, and so on). They trained the model using a 10-fold cross-validation for both the PLS and the SVR models. For model optimization they divided the dataset in approximately equal subsets, were 9 were used to construct the model, the 10th subset was used for validation purposes. Prediction accuracy of these models was evaluated using the Pearson’s correlation coefficient. They observed that the PLS models resulted in slightly better predictive capabilities. However, when input data were rather noisy the SVR models were more robust in comparison with the PLS models. Therefore they presented the final results based on the SVR models. As expected the cumulative data sets showed the highest Pearson’s correlation coefficients, followed by the production scale datasets (12,000 L) and the small scale datasets (80 L) showed the lowest correlation coefficients, see Figure 6. In this figure the top 20 % are given in blue while the bottom 20 % is given in red, a-d shows the actual titer versus the predicted titer of each case: (a) 80 L scale, (b) 2000 L scale, (c) up to 70 h of 12,000 L scale and (d) up to 260 h of 12,000 L scale. In Figure 6 the progression of the predicted titer is shown over the course of the cultures for the top and the bottom 20% runs.

S. Sangers

20

Figure 6: SVR models’ prediction accuracy of the final titer using different datasets. The correlation coefficient (r) between the predicted and the actual titer is shown for each case. The dashed lines indicate the separation of the top 20%, middle 60%, and bottom 20% of runs based on the predicted titer (y-axis) or the actual titer (x-axis). Runs in the top 20% class based on the actual titer are colored in blue; runs in the middle 20% class are colored in gray; and runs in the bottom 20% class are colored in red. (a) 80 L scale. (b) 2000 L scale. (c) Up to 70 h of 12,000 L scale. (d) Up to 260 h of 12,000 L scale. (e) The progression of predicted titer is shown over the course of the cultures for the top and the bottom 20% runs. 11

Nevertheless, it was found that within the small scale data sets, from the top 20% runs only a few runs actually fell to the bottom 20%. At 70 h into the production no class switching was occurring at all, indicating that the process characteristics at the early stage of production are indicative for the final outcome. Thus intervention should take place prior 70 h into the production process in order to influence the outcome.

Both Streefland et al. and Le et al. showed the potential benefits of implementing a QbD based production process. Streefland et al. clearly showed how to apply all the steps required with a rather limited dataset and they were able to predict the product quality with their constructed PCA model. Getting close to actual release the product based on real-time data. Sadly, insufficient data was available and their setup should still be operating next to the normal quality control testing procedures. By using a larger data set Le et al. showed that significant more information could be retrieved from the process data. They constructed a SVR model with predictive capabilities with regard to the lactate and final

S. Sangers

21

process are decisive for the final product titer.

7

Discussion

The FDA published the PAT guidance framework for the industry,3 as they acknowledged the potential benefits for (bio)pharmaceutical manufacturers are tremendous. With this guidance they tried to encourage the biopharmaceutical manufacturers to implement QbD and PAT for their production processes. This PAT approach was initially applied within the API manufacturing.14 During the last decades, biotech manufactures have made significant advances to implement QbD and PAT on their complex production processes.4,45,57,58 These QbD and PAT setups have led to increased understanding of the production process, this could eventually lead to actually build the quality into the manufacturing process. In order to monitor the quality of a product, the CQA’s are to be analyzed. Ideally the CQA’s are directly monitored by PAT. However, direct analysis of these attributes is often an analytical challenge, and therefore correlated CPP’s are analyzed instead.14,34 First the CQA’s and CPP’s for a specific production process have to be identified, ideally this is done by a DOE setup, followed up by MVDA. However early process development is rather unstructured, but does contain relevant process information, which is traditionally also used for determination of CQA’s and CPP’s. These data can still be used to gather additional process information using MVDA, as Mercier et al. reported,43 where they truncated data so that all data sets were of similar dimensions. Alternatively they could have applied other strategies to account for the incomplete data sets. For instance maximum likelihood principal component analysis (MLPCA), could have been used as described by Andrews et al.59. Or one of the PCA models described by Folch-Fortuny et al.,60 by using one of these approaches it would have been possible to proceed with the all the data instead of truncating parts of it.

Next a multivariate model of the production process should be constructed, by using this model the process can be monitored real time with the appropriate PAT sensors for the CPP’s or directly the CQA’s as previous determined. These models should be well trained and validated in order to make them capable to predict when the product process will go out of spec, and allowing for real time intervention of the process to be corrected. Therefore, biopharmaceutical manufacturers should strive to develop the capability to monitor the CPP’s and the CQA’s and understand the process, thereby Quality can be

S. Sangers

22

pharmaceutical manufacturers are conservative and very reluctant to actually implement PAT as recommended by the FDA. Even stating that the potential benefits of PAT are not routinely investigated due to personnel and budget cuts.8 Rathore made similar conclusions in 2014, while QbD implementation has taken a leap the past decade, actual PAT adoption is still lacking within the biotech production processes.4 One of the reasons could be that the industry perception of the PAT benefits versus the costs of doing so and the resulting risks and complexity isn’t worth their effort. Resulting in only a few QbD processes are approved by the regulators.4,6 On the other hand there seems to be collaboration between the industry and the regulators in order to reach agreement on the QbD principles and filing requirements.6

In addition, within the field there are numerous examples of researchers who have reported PAT and QbD implementations for biotech production processes. For instance Streefland et al.,10 demonstrated how the general principals of PAT and design space can be applied for monitoring and controlling a cultivation process of a biological product. Using their PCA model they were able to assure in real time that the process is running within the process design space. While Le et al.11 reported SVR and PLS model capable to predict final product titer and lactate concentration in a cell culture process. These models were constructed using process data of 134 parameters from 243 production runs. When evaluating the models, they found that the first 70 h into the production process are decisive for the product titer, and thus remediation’s should be performed prior this time point.

So is the biopharmaceutical industry currently capable for real time release of their products, as the QbD approach was promising all along. A recent survey of Rita C. Peters showed that QbD has improved process understanding and product quality, however 32 % of the respondents had not implemented QbD, with as main reason the lack of guidance and direction from regulatory agencies.61 _{It seems that}

this is the final struggle the industry has to make before full implementation of QbD and PAT is realized. And indeed as far my knowledge, there are no publications regarding a real-time releasing of products from a biotech production process. This might also due to the complexity of properly interpreting multivariate data to establish the product quality. Traditionally industry and agencies have been applying QbT, and thus the end products are release tested on certain CQA’s. It will take significant steps

S. Sangers

23

challenges as well.

So in order to establish real-time release testing it might be a good strategy to analyze the CQA’s directly by utilizing at-line PAT. Although this strategy might require additional analytical improvements, it seems to be more feasible in comparison with MVDA models based on PAT data. For one, the filings can still be based on a similar format as used nowadays within the QbT paradigm. Instead of the new format required for MVDA based control processes, which require more abstract performance indicators. In addition, the data from the online sensors in the PAT setup can be analyzed with MVDA analysis to further enlarge process understanding. At a later stage this process knowledge in can be used to construct MVDA models which are capable of monitoring and controlling the process using more abstract performance indicators. Finally, proper understanding of the complex multivariate data sets from the biopharmaceutical production processes could enable real time release of biopharmaceuticals.

To conclude, during the last decades, the biotech industry has made significant progress with regard to knowledge of their production processes by utilizing PAT. By combining the data from PAT and reference methods and subsequently applying MVDA, increased process understanding was obtained. This in-depth process knowledge is essential for proper implementation of QbD and PAT. Many papers are available of research groups applying this process on mostly pilot scales. There are some reports of biotech products being manufactured under QbD.4,6 However, there are, to the best of our knowledge, no reports of real-time releasing of biotech products by using QbD. Therefore it is concluded that product quality determination by the utilizing QbD and PAT is not yet the typical practice within the biopharmaceutical industry and the QbT remains the standard way of manufacturing biopharmaceutical products.

8

References

1. Yu, L. X. Pharmaceutical quality by design: Product and process development, understanding, and control. Pharm. Res. 25, 781–791 (2008).

2. Del Val, I. J., Kontoravdi, C. & Nagy, J. M. Towards the implementation of quality by design to the production of therapeutic monoclonal antibodies with desired glycosylation patterns. Biotechnol.

S. Sangers

24

Manufacturing, and Quality Assurance. 16 (2004).

4. Rathore, A. S. QbD/PAT for bioprocessing: Moving from theory to implementation. Curr. Opin.

Chem. Eng. 6, 1–8 (2014).

5. Rathore, A. S. & Winkle, H. Quality by design for biopharmaceuticals. Nat. Biotechnol. 27, 26–34 (2009).

6. Luciani, F., Galluzzo, S., Gaggioli, A., Kruse, N. A., Venneugues, P., Schneider, C. K., Pini, C. & Melchiorri, D. Implementing quality by design for biotech products: Are regulators on track?

MAbs 7, 451–455 (2015).

7. Sagmeister, P., Wechselberger, P., Jazini, M., Meitz, A., Langemann, T. & Herwig, C. Soft sensor assisted dynamic bioprocess control: Efficient tools for bioprocess development. Chem. Eng. Sci.

96, 190–198 (2013).

8. Sonnleitner, B. Automated Measurement and Monitoring of Bioprocesses: Key Elements of the M3C Strategy. Adv. Biochem. Eng. Biotechnol. 123, 1–33 (2012).

9. Hakemeyer, C., Strauss, U., Werz, S., Jose, G. E., Folque, F. & Menezes, J. C. At-line NIR spectroscopy as effective PAT monitoring technique in Mab cultivations during process development and manufacturing. Talanta 90, 12–21 (2012).

10. Streefland, M., Van Herpen, P. F. G., Van De Waterbeemd, B., Van Der Pol, L. A., Beuvery, E. C., Tramper, J., Martens, D. E. & Toft, M. A practical approach for exploration and modeling of the design space of a bacterial vaccine cultivation process. Biotechnol. Bioeng. 104, 492–504 (2009). 11. Le, H., Kabbur, S., Pollastrini, L., Sun, Z., Mills, K., Johnson, K., Karypis, G. & Hu, W. S. Multivariate

analysis of cell culture bioprocess data-Lactate consumption as process indicator. J. Biotechnol.

162, 210–223 (2012).

12. Scott, B. Process Analytical Technology in the Pharmaceutical Industry : A Toolkit for Continuous Improvement. 60, (2006).

13. Bordawekar, S., Chanda, A., Daly, A. M., Garrett, A. W., Higgins, J. P., LaPack, M. A., Maloney, T. D., Morgado, J., Mukherjee, S., Orr, J. D., Reid, G. L., Yang, B. S. & Ward, H. W. Industry Perspectives on Process Analytical Technology: Tools and Applications in API Manufacturing. Org.

Process Res. Dev. 19, 1174–1185 (2015).

14. Rathore, A. S., Read, E. K., Park, J. T., Shah, R. B., Riley, B. S. & Brorson, K. A. Process Analytical Technology (PAT) for Biopharmaceutical Products: Part I. Concepts and Applications. Anal.

Bioanal. Chem. 398, 137–154 (2010).

15. Nelson, A. L., Dhimolea, E. & Reichert, J. M. Development trends for human monoclonal antibody therapeutics. Nat. Rev. Drug Discov. 9, 767–774 (2010).

16. Warnock, J. N. & Al-Rubeai, M. Bioreactor systems for the production of biopharmaceuticals from animal cells. Biotechnol. Appl. Biochem. 45, 1–12 (2006).

17. Jain, E. & Kumar, A. Upstream processes in antibody production: Evaluation of critical parameters. Biotechnol. Adv. 26, 46–72 (2008).

S. Sangers

25

antibodies-Application of platform approaches. J. Chromatogr. B Anal. Technol. Biomed. Life Sci.

848, 28–39 (2007).

19. Gronemeyer, P., Ditz, R. & Strube, J. Trends in Upstream and Downstream Process Development for Antibody Manufacturing. Bioengineering 1, 188–212 (2014).

20. Junker, B. H. & Wang, H. Y. Bioprocess monitoring and computer control: Key roots of the current PAT initiative. Biotechnol. Bioeng. 95, 226–261 (2006).

21. Teixeira, A. P., Oliveira, R., Alves, P. M. & Carrondo, M. J. T. Advances in on-line monitoring and control of mammalian cell cultures: Supporting the PAT initiative. Biotechnol. Adv. 27, 726–732 (2009).

22. Simon, L. L. et al. Assessment of Recent Process Analytical Technology (PAT) Trends: A Multiauthor Review. (2014).

23. Mandenius, C. & Brundin, a. REVIEW : BIOCATALYSTS AND BIOREACTOR DESIGN Optimization, Bioprocess Methodology, Using Design-of-experiments. Biotechnol Progr 24, 1191–1203 (2008). 24. RATHORE, A. S., SHARMA, A. & CHILIN, D. Applying process analytical technology to biotech unit

operations. Biopharm Int. 19,

25. Hou, Y., Jiang, C., Shukla, A. A. & Cramer, S. M. Improved process analytical technology for protein a chromatography using predictive principal component analysis tools. Biotechnol.

Bioeng. 108, 59–68 (2011).

26. Rathore, A. S., Mittal, S., Lute, S. & Brorson, K. Chemometrics applications in biotechnology processes: Predicting column integrity and impurity clearance during reuse of chromatography resin. Biotechnol. Prog. 28, 1308–1314 (2012).

27. Jacquemart, R., Chavane, N., Durocher, Y., Hoemann, C., De Crescenzo, G. & Jolicoeur, M. At-line monitoring of bioreactor protein production by surface plasmon resonance. Biotechnol. Bioeng.

100, 184–188 (2008).

28. Larson, T. M., Gawlitzek, M., Evans, H., Albers, U. & Cacia, J. Chemometric evaluation of on-line high-pressure liquid chromatography in mammalian cell cultures: Analysis of amino acids and glucose. Biotechnol. Bioeng. 77, 553–563 (2002).

29. Rathore, A. S., Parr, L., Dermawan, S., Lawson, K. & Lu, Y. Large scale demonstration of a process analytical technology application in bioprocessing: Use of on-line high performance liquid chromatography for making real time pooling decisions for process chromatography. Biotechnol.

Prog. 26, 448–457 (2010).

30. Carvell, J. P. & Dowd, J. E. On-line measurements and control of viable cell density in cell culture manufacturing processes using radio-frequency impedance. Cytotechnology 50, 35–48 (2006). 31. Joeris, K., Frerichs, J. G., Konstantinov, K. & Scheper, T. In-situ microscopy: Online process

monitoring of mammalian cell cultures. Cytotechnology 38, 129–134 (2002).

32. Kacmar, J. & Srienc, F. Dynamics of single cell property distributions in Chinese hamster ovary cell cultures monitored and controlled with automated flow cytometry. J. Biotechnol. 120, 410–420

S. Sangers

26

33. Teixeira, A. P., Portugal, C. A. M., Carinhas, N., Dias, J. M. L., Crespo, J. P., Alves, P. M., Carrondo, M. J. T. & Oliveira, R. In situ 2D fluorometry and chemometric monitoring of mammalian cell cultures. Biotechnol. Bioeng. 102, 1098–1106 (2009).

34. Pais, D. A. M., Carrondo, M. J. T., Alves, P. M. & Teixeira, A. P. Towards real-time monitoring of therapeutic protein quality in mammalian cell processes. Curr. Opin. Biotechnol. 30, 161–167 (2014).

35. Doherty, M., Bones, J., McLoughlin, N., Telford, J. E., Harmon, B., DeFelippis, M. R. & Rudd, P. M. An automated robotic platform for rapid profiling oligosaccharide analysis of monoclonal antibodies directly from cell culture. Anal. Biochem. 442, 10–18 (2013).

36. Burnina, I., Hoyt, E., Lynaugh, H., Li, H. & Gong, B. A cost-effective plate-based sample preparation for antibody N-glycan analysis. J. Chromatogr. A 1307, 201–206 (2013).

37. Beck, A., Diemer, H., Ayoub, D., Debaene, F., Wagner-Rousset, E., Carapito, C., Van Dorsselaer, A. & Sanglier-Cianférani, S. Analytical characterization of biosimilar antibodies and Fc-fusion proteins. TrAC Trends Anal. Chem. 48, 81–95 (2013).

38. Amand, M. M. S., Ogunnaike, B. A. & Robinson, A. S. Development of at-line assay to monitor charge variants of MAbs during production. Biotechnol. Prog. 30, 249–255 (2014).

39. Konno, Y., Kobayashi, Y., Takahashi, K., Takahashi, E., Sakae, S., Wakitani, M., Yamano, K., Suzawa, T., Yano, K., Ohta, T., Koike, M., Wakamatsu, K. & Hosoi, S. Fucose content of monoclonal antibodies can be controlled by culture medium osmolality for high antibody-dependent cellular cytotoxicity. Cytotechnology 64, 249–265 (2012).

40. Grainger, R. K. & James, D. C. CHO cell line specific prediction and control of recombinant monoclonal antibody n-glycosylation. Biotechnol. Bioeng. 110, 2970–2983 (2013).

41. Mercier, S. M., Diepenbroek, B., Wijffels, R. H. & Streefland, M. Multivariate PAT solutions for biopharmaceutical cultivation: Current progress and limitations. Trends Biotechnol. 32, 329–336 (2014).

42. Kourti, T. The Process Analytical Technology initiative and multivariate process analysis, monitoring and control. Anal. Bioanal. Chem. 384, 1043–1048 (2006).

43. Mercier, S. M., Diepenbroek, B., Dalm, M. C. F., Wijffels, R. H. & Streefland, M. Multivariate data analysis as a PAT tool for early bioprocess development data. J. Biotechnol. 167, 262–270 (2013). 44. Rathore, A. S., Read, E. K., Park, J. T., Shah, R. B., Riley, B. S. & Brorson, K. A. Process Analytical

Technology (PAT) for Biopharmaceutical Products: Part II. Concepts and Applications. Anal.

Bioanal. Chem. 398, 137–154 (2010).

45. Glassey, J., Gernaey, K. V., Clemens, C., Schulz, T. W., Oliveira, R., Striedner, G. & Mandenius, C. F. Process analytical technology (PAT) for biopharmaceuticals. Biotechnol. J. 6, 369–377 (2011). 46. Rantanen, J. & Khinast, J. The Future of Pharmaceutical Manufacturing Sciences. J. Pharm. Sci.

S. Sangers

27

identification and successful resolution of a root cause for a bioprocessing application. TL - 24.

Biotechnol. Prog. 24 VN - r, 720–726 (2008).

48. Carinhas, N., Oliveira, R., Alves, P. M., Carrondo, M. J. T. & Teixeira, A. P. Systems biotechnology of animal cells: The road to prediction. Trends Biotechnol. 30, 377–385 (2012).

49. Kitano, H. Computational systems biology. Nature 420, 206–210 (2002).

50. Clavaud, M., Roggo, Y., Von Daeniken, R., Liebler, A. & Schwabe, J. O. Chemometrics and in-line near infrared spectroscopic monitoring of a biopharmaceutical Chinese hamster ovary cell culture: Prediction of multiple cultivation variables. Talanta 111, 28–38 (2013).

51. Kirdar, A. O., Conner, J. S., Baclaski, J. & Rathore, A. S. Application of multivariate analysis toward biotech processes: Case study of a cell-culture unit operation. Biotechnol. Prog. 23, 61–67 (2007). 52. Gunther, J. C., Conner, J. S. & Seborg, D. E. Fault detection and diagnosis in an industrial

fed-batch cell culture process. Biotechnol. Prog. 23, 851–857 (2007).

53. Nucci, E. R., Cruz, A. J. G. & Giordano, R. C. Monitoring bioreactors using principal component analysis: production of penicillin G acylase as a case study. Bioprocess Biosyst. Eng. 33, 557–64 (2010).

54. Navratil, M., Norberg, A., Lembren, L. & Mandenius, C. F. On-line multi-analyzer monitoring of biomass, glucose and acetate for growth rate control of a Vibrio cholerae fed-batch cultivation. J.

Biotechnol. 115, 67–79 (2005).

55. Streefland, M., van de Waterbeemd, B., Happ??, H., van der Pol, L. A., Beuvery, E. C., Tramper, J. & Martens, D. E. PAT for vaccines: The first stage of PAT implementation for development of a well-defined whole-cell vaccine against whooping cough disease. Vaccine 25, 2994–3000 (2007). 56. Wold, S. K. N. F. H. & Holmberg, A. Modelling and Diagnostics of Batch Processes and Analogous

Kinetic Experiments. 331–340 (1998).

57. Alford, J. S. Bioprocess control: Advances and challenges. Comput. Chem. Eng. 30, 1464–1475 (2006).

58. Bhambure, R., Kumar, K. & Rathore, A. S. High-throughput process development for biopharmaceutical drug substances. Trends Biotechnol. 29, 127–135 (2011).

59. Andrews, D. T. & Wentzell, P. D. Applications of maximum likelihood principal component analysis: Incomplete data sets and calibration transfer. Anal. Chim. Acta 350, 341–352 (1997). 60. Folch-Fortuny, A., Arteaga, F. & Ferrer, A. PCA model building with missing data: New proposals

and a comparative study. Chemom. Intell. Lab. Syst. 146, 77–88 (2015).

61. Peters, R. C. 2014 Manufacturing Trends and Outlook. Available at: http://www.biopharminternational.com/2014-manufacturing-trends-and-outlook. (Accessed: 19th June 2016)