Abstract
The imperative to reduce process development times and costs, as well as batch-to-batch variation is known across the biopharmaceutical industry; ultimately helping to ensure earlier patient treatment and reduced biopharmaceutical prices. With the Process Analytic Technology (PAT) initiative the regulatory authorities have provided an opportunity to address those topics, harnessing novel risk and scientific approaches, in which process understanding is the central element. Industry has started to adopt such approaches, e.g. the Quality by Design (QbD) paradigm, and is also not shy to invest in digitalization and Industry 4.0. However, the implementation of a comprehensive and conclusive concept for knowledge management still seems largely missing. Two main issues are the inability of exploiting available mechanistic understanding and the attempt to understand dynamic processes with static experiments. Addressing these issues would allow reducing process development timelines and costs and also enable the application of advanced Model Predictive Control (MPC) attacking batch-to-batch variations, as in this article demonstrated for a bioreactor process.
Issues in Biopharmaceutical Drug Development
Biopharmaceutical drug development is an ambivalent affair. On the one side, the demand for innovative drugs and the revenue generated by existing biopharmaceuticals is growing reaching $200 billion per year soon. Contrary the development costs for a new a drug is around $1.8 billion while the development time still being around 13.5 years.1 High dropout rates in clinical trials, the long development times for new biopharmaceuticals entering the market and the resulting short remaining patent protection burden both companies and customers. Hence a huge pressure is applied by authorities to shorten development times and costs. Due to reduced clinical testing, cost efficiency and short developing times are even more crucial for biosimilar manufacturers. Since an altered number of new companies are into new drug and biosimilar development, the demand for shorter process development times is also strongly market driven.
Looking at the whole production chain, the required time for fermentation development, optimization and characterization appears to be a major cost driver. Compared to other unit operations like filtration or chromatography, cell cultivation requires at least one order of magnitude longer. Therefore, reducing the time required to sufficiently describe the fermentation process in a faster way, would shorten time to market and significantly reduce costs.
A major part of the puzzle to shrink development effort was presented by the US Food and Drug Administration (FDA) over a decade ago by the PAT and QbD guidance initiatives. The sound understanding of the sources and implication of variability in both process and product is the core principle of PAT/QbD. Coming from the target product prof le, the critical quality attributes (CQAs) are determined and are subsequently used to identify potential critical control parameters (CCPs) via risk assessment. Experimental studies are then used to evaluate which and how changes in the CPPs impact on the product quality (CCPs -> CQAs). To develop understanding of the process, Design of Experiment (DoE) is the standard method applied in bioprocess development because it allows investigating the impact of changes in the CPPs on the CQAs in an optimal statistical fashion. DoE also helps to characterize the design space and possibly reduces development time by using an educated setup of experimental points. However, the commonly used DoE approaches have a major Achilles tendon: they require the process parameters to be static (constant) during the process. Changes in the process parameters (e.g. temporal deviation) during the process are not considered. Hence, also slight changes in CCPs during the process cannot be described by statistical models that are developed on those data. The dynamics of the process are not covered and approaches that could counteract on temporal deviations, such as model predictive control, cannot be set up on the statistical models because they require a dynamic description of the process. With advanced control approaches in hand variations in the process could be levelled out by manipulating the CCPs such that a constant product quality is achieved (Figure 1). The modeling of process dynamics and their consideration during experimental design, does not increase, but to the contrary significantly decreases the number of experiments that have to be carried out, allowing for a clear reduction in the overall development time. In what follows, it is shown how an intensified DoE method for the investigation of the process dynamics and the description of the impact of CCPs on CQAs by a reliable dynamic model can be combined with real time monitoring of the process state, to enable model predictive control, while at the same time increasing the process understanding and decreasing the number of experiments, as compared to the standard approach.
Figure 1. (A): Comparison between a fixed bioprocess and a process applying model predictive control to ensure constant quality output.Enablers
Multivariate Data Analysis (MVDA) and Soft Sensors
One key aspect to be able to obtain constant product quality is to investigate cause and effect between input and output. It was already demonstrated several times how MVDA, such as principle component analysis, partial least square analysis, random forest or neural networks can unravel important correlations improving the process understanding and adding to the milestone of QbD.2-5 However, the application of MVDA models based on data generated by DoE is limited to the region explored by the DoE as long as no general mechanistic understanding is established. Like for most statistical methods MVDA does not extrapolate well, wherefore the results need to be considered carefully when using the algorithm on data it was not trained on. Hence, transferring the gathered knowledge to different scales or processes might be challenging or even not feasible.
MVDA can be also be used to design soft sensors by correlating high frequent online data from sensors like fl uorescence, infrared or even standard process data with offl ine measured CQAs. After model calibration it is thereby possible to monitor product quality online. This approach is very valuable for process understanding and monitoring. Since sensor data do not impact on the process, but only monitor the state, a MVDA approach based on spectroscopic data cannot be used for process control, i.e. the operator cannot change sensor data to control product quality. Therefore, models that are solely based on CCPs that can be used for manipulating the process are the key to success when aiming at MPC. All in all, MVDA seems to be a very promising tool for process monitoring but for application of a MPC other approaches are required.
Hybrid Modeling
The counterpart to MVDA is fundamental modeling (such as mechanistic or first-principle modeling) were the mechanism of the cause and effect for a certain system are described. However, describing fermentation processes with fundamental modeling approaches is difficult, requires experience and takes a considerable amount of time due to the large number of unknown interferences in fermentations and the enormous complexity. An alternative to pure statistical or fundamental methods can be found in hybrid models. Hybrid models combine process knowledge (also called white box models) like mechanistic models with MVDA (black box models).
For fermentations, the modeling of time dependant variables like the specific growth rate µ, can be accomplished by combining e.g. a Monod model that describes the impact of substrate concentration (S) with a black box that captures the impact of temperature and pH: µ= µmax(Temp, pH)*S/(K+S), with K as constant. As can be seen from this example, variations in the substrate concentration (feeding rate) could be compensated by changing other critical parameters such as pH or process temperature to keep the specific growth rate constant, as also described elsewhere.6 It is relatively easy to develop a hybrid model and in fact the dynamic material balance provide a generally valid backbone to the hybrid model development. The relative ease of development, the dynamic process description and the possibility to describe the impact of changing CCPs on CQAs makes hybrid models an ideal candidate for MPC. In addition, and due to the mechanistic backbone, it is easier to transfer the model when scale up and down or to re-use it in new projects, wherefore it helps with knowledge management. Thus, using hybrid models together with DoE to identify the impact of the inputs is already a major step towards MPC. However, the dynamic domain is still not captured in the data due to the rigid endpoint contemplation of each experiment.
Intensified Design of Experiment (iDoE)
In order to uncover the dynamics of processes, intensified DoE (iDoE) is one promising candidate. Here dynamics are unleashed by changing CCPs within an experiment, e.g. cell cultivation several times. Hence, the response of the process to slight temporal changes within the process can directly be addressed. This method thereby grants two substantial benefits; description of dynamics and time savings. While the dynamic can be assessed by changing the input parameters a number of times during the processes, e.g. three to four times during each fermentation, the results of each change can be seen as the endpoint of a static point in a classical DoE. Although some points within the design space have to be tested from different sides,7 a comparable DoE design space can be described with iDoE two to three times faster. However, the iDoE method also requires an increased sampling and a model being able to describe the dynamic of the system. Hybrid modeling is an effective tool for obtaining a dynamic model, as highlighted before. Hence, the combination of a hybrid model together with iDoE enables access to the dynamic domain, gaining understanding about the impact of temporal process disturbances. The combination further reduces the number of required experiments to describe a certain design space significantly, while also enabling MPC.
Data Management and Software Requirements
To be able to control product quality online and in real time the process environment must collect data, feed the identified critical variables to the model algorithm, simulate complex models in real time and give a feedback which CCP has to be changed in order to maintain CQAs. This is a complex task and many current automation systems are not able to cover the requirements due to following facts. Many current distributed control systems (DCS) used for automation in the biopharmaceutical environment possess a limited computational capability of the applied microcontroller.8 Hence, complex tasks like applying highly non- linear MPC to a certain process is unfeasible. MPC also requires fully automated real time data pre-processing (filtering, smoothing, outlier detection), model calculation as basis to take process control actions. Other superior systems are required to fulfill such work like a supervised control and data acquisition (SCADA) or advanced data management system. Here both the computational power is provided to cope with complex optimization routines and further advanced sensors (FT-IR, Fluorescence, PTR-MS) and moreover soft-sensor codes can be integrated. The flexible environment in process development (combination of sensors and reactors and PAT analysers) demands a flexible software environment. The idea of introducing a workflow based approach in industrial R&D facilities shows promising results for standardized, scalable and modular data management systems. However, the integration of such systems to an existing system requires clear specifications for the vendor and several management decisions requiring time and strategic thinking.
What is the Missing Link?
The combination of hybrid modeling and intensified DoE can be seen as key enablers for modeling predictive control, as already described. Currently powerful industrial software systems with the computational power to deal with complex models are finding their way into industries. However, there is currently no commercially available hybrid modeling toolbox or software to exploit iDoEs on the market. The ideal workflow would be to generate the design space with iDoE, develop a hybrid model containing CCPs to control CQAs and subsequently exploit the model for reliable MPC (Figure 2). This would mean that once a certain hybrid model is established, the parameters of the code have to be transferred to the SCADA system. While collecting data from the DCS, the SCADA has to take decisions based on the data and the model, returning execution commands back to the DCS system. Such a workflow would enable model predictive control for bioprocesses. Currently, many of the described parts are already available, while some are still under construction. The outline of the puzzle for model predictive control is visible and within the Novasign project a hybrid model- and intensified DoE toolbox is currently under implementation.
Figure 2. Schematic description of a model predictive control process for cell cultivation. An iDoE- and a hybrid model toolbox are applied for design space characterization and model identification, respectively. The SCADA systems applies the process data onto the identified model, takes decisions and transfers the required information for CCP changes to the DCS for execution.References
- Paul, S.M., D.S. Mytelka, C.T. Dunwiddie, C.C. Persinger, B.H. Munos, S.R. Lindborg, and A.L. Schacht, How to improve RD productivity: The pharmaceutical industry’s grand challenge. Nature Reviews Drug Discovery, 2010. 9(3): p. 203-214.
- Kirdar, A.O., K.D. Green and A.S. Rathore, Application of multivariate data analysis for identification and successful resolution of a root cause for a bioprocessing application. Biotechnology Progress, 2008. 24(3): p. 720-726.
- Ündey, C., S. Ertun, T. Mistretta and B. Looze, Applied advanced process analytics in biopharmaceutical manufacturing: Challenges and prospects in real-time monitoring and control. Journal of Process Control, 2010. 20(9): p. 1009-1018.
- Melcher, M., T. Scharl, B. Spangl, M. Luchner, M. Cserjan, K. Bayer, F. Leisch, and G. Striedner, The potential of random forest and neural networks for biomass and recombinant protein modeling in Escherichia coli fed-batch fermentations. Biotechnology Journal, 2015. 10(11): p. 1770-1782.
- Sokolov, M., J. Ritscher, N. MacKinnon, J. Souquet, H. Broly, M. Morbidelli, and A. Butté, Enhanced process understanding and multivariate prediction of the relationship between cell culture process and monoclonal antibody quality. Biotechnology Progress, 2017. 33(5): p. 1368-1380.
- von Stosch, M., J.M. Hamelink and R. Oliveira, Hybrid modeling as a QbD/PAT tool in process development: an industrial E. coli case study. Bioprocess and Biosystems Engineering, 2016. 39(5): p. 773-784.
- von Stosch, M., J.M. Hamelink and R. Oliveira, Toward intensifying design of experiments in upstream bioprocess development: An industrial Escherichia coli feasibility study. Biotechnology Progress, 2016. 32(5): p. 1343-1352.
- Sommeregger, W., B. Sissolak, K. Kandra, M. von Stosch, M. Mayer, and G. Striedner, Quality by control: Towards model predictive control of mammalian cell culture bioprocesses. Biotechnology Journal, 2017. 12(7).