Biopharmaceuticals represent a significant and growing sector of the pharmaceutical industry. The global biopharmaceutical industry is currently worth over €107 billion, according to research conducted by BioPlan Associates (1). The industry should exceed €123 billion in 2015, says the International Market Analysis Research and Consulting Group (2). Biopharmaceuticals are used in the prevention and treatment of disease and there are over 300 approved biopharmaceuticals on the market, with many in clinical development yet to come. As with all pharmaceutical compounds, the cost and time required to bring a biopharmaceutical product to market has huge economic implications. The average cost of researching and developing a new medicine is claimed by the pharmaceutical industry to be approximately €1.4 billion (3). Process development for biopharmaceutical products is generally more labour intensive, time consuming and expensive than for comparable traditional pharmaceutical processes because of the large number of individual processes and potential variables involved. In recent years, there has been a growing necessity to increase the speed, efficacy and information content associated to the development and scale-up of biopharmaceutical processes, particularly given the regulatory authorities current shift in philosophy towards quality-by-design (QbD). This QbD requirement for increased process understanding intensifies the work involved in the earlier phases of the drug-product lifecycle. The benefit, however, is that the increased process knowledge can speed the technical transfer from development into manufacturing, deliver a more optimised, robust process with higher titres and greater reproducibility and aid in troubleshooting and root-cause analysis of deviations during production. Taking these factors into consideration, evolving a process development strategy that reduces costs and timelines while simultaneously providing greater process knowledge would be highly advantageous. CURRENT STATE OF PLAY Initial bioprocess development consists of the parallel strands of cell-line optimisation, clone selectio, and screening for media, feed components and strategies, as well as other process conditions. [caption id="attachment_15960" align="alignright" width="1024"] Figure 1: Bioprocess development streams (click to enlarge)[/caption] Shake flasks, the most common vessels used in early cell work, have served the biopharmaceutical industry well over the decades, but their limitations for optimising cell culture conditions are well known. Shake flasks allow control of temperature, ambient gas mix and agitation rate, but standard upstream bioprocess monitoring and controlling of critical process parameters such as pH, dissolved oxygen (DO) and feed schedules are beyond the capabilities of these vessels. However, these critical process parameters influence cell metabolism, viability and productivity, and ultimately product quality and stability. Operation within such broad design-space variability intensifies the difficult task facing bioprocess developers in identifying the superiority of one clone over another or the influences of media, feed and supplementation strategies – factors that are instrumental in improving volumetric productivity. Selecting suboptimal clones during early development when using shake flasks is not uncommon and diminished cell productivity and product quality then persist through development and beyond. Bioprocess development platforms used during early-stage process development should mimic the physical and mechanical characteristics of production-scale reactors to the greatest degree possible, to ensure consistency throughout development phases. Bench-top bioreactors have the potential to address process consistency and harmonise unit operations between development and production. [caption id="attachment_15962" align="alignright" width="1024"] Figure 2: Operating the DASGIP bench-top bioreactor workstation[/caption] However, traditionally bioreactors only routinely monitor pH, temperature and dissolved oxygen (DO) online. Unfortunately, these routinely measured variables do not provide significant insight into the mechanisms of the process. The lack of routine in-line measurements of critical process parameters (CPPs) reduces the potential for increased process understanding and ultimately the potential for direct control of the critical quality attributes (CQAs) of the product by manipulation of the appropriate CPPs. The level of process understanding that can or should be achieved beyond the acceptable minimum level promises to be the scope of a continuing debate among the biopharmaceutical industry and its regulators. In practice, the path of increased understanding may follow a series of incremental steps toward the desired, optimal bioprocess design space. Development of such understanding beyond information collected from product and process characterisation studies during development can come from using a process analytical technology (PAT) approach for process monitoring. PAT-ENABLED PROCESS DEVELOPMENT PLATFORM In 2004, the United States Food and Drug Administration (FDA) defined process analytical technology (PAT) as a mechanism to design, analyse, and control pharmaceutical manufacturing processes through the measurement of CPPs which affect CQAs (4). The philosophy behind the FDA PAT initiative is that the CQAs of a product are directly determined by the CPPs. [caption id="attachment_16087" align="alignright" width="800"]New Picture Figure 3: QbD philosophy for process development[/caption] Therefore, the delivery of the desired CQAs can be ensured if the CPPs are identified, the nature of their relationships to the CQAs understood and appropriate control strategies then applied to guarantee a high quality, reproducible output from the process. Overall, this QbD philosophy states that designing a robust process depends on the interplay of two distinct factors: the level of process understanding achieved and the level of process control implemented (Figure 3, right). Processes that are both reproducible and robust can be achieved only with high levels of process understanding and process control. Optimal cell growth is achieved only through a narrow range of environmental conditions. It is evident from Figure 3 that it is possible to operate a reproducible bioreactor process within narrow operating ranges without a high level of process understanding. However, bioreactor control provides special challenges due to significant process variability, the complexity and nonlinearity of biological systems, the need to operate in a sterile environment, and the relatively few real-time direct measurements available that help define the state of the culture. Bioprocesses employ most of the same types of control as are used in other chemical industries,  which consist mainly of traditional single input single output feedback PI (proportional + integral) controllers. These simple controllers are used to control the bioprocess variables (pH, DO, temperature) which are measured routinely online at regular sample intervals. However, these control algorithms do not take into account the dynamics of the process and thus act solely to reduce the error between a defined set-point and the process variable of interest. Incorporating process knowledge into a control algorithm can enhance robustness and direct the process along the optimal batch trajectory. Processes that are both reproducible and robust can be achieved only with high levels of process understanding and process control. So it is not surprising that the PAT guidance emphasises both factors. Measurement: Acquisition of Process Data There is a growing interest in the concepts of product and process optimisation based on PAT to ensure public safety and product efficacy. The positive impacts of PAT that have been seen in small-molecule drug development and manufacturing are a significant part of the reason that the concepts are now being applied to the biopharmaceutical sector. PAT is enabling a more fundamental understanding of how bioprocesses work and what influences their efficiency. There are many complexities associated with the cell-line chosen, the media, the quality of the nutrients, the processing conditions and the harvesting of the product. PAT can be used comprehensively in bioprocessing to help manufacturers exercise greater control over operations and simplify some of the complexity. In-line process variables that could not be monitored in the past can now be measured, analysed and used for advanced control schemes. Process analytical technologies act as the ‘eyes’ inside the bioreactor. In bioprocesses, the need for real-time process information is particularly high, due to the complexity and unpredictability of the process. The media used to support growth and protein production in animal culture are extremely sophisticated mixtures, often containing in excess of eighty different species, almost all of which are at very low concentrations when compared with small molecule production. Many of the species are from similar families, such as the twenty or so amino acids used by mammalian cells. The low concentrations combined with the structural similarities of multiple species means that finding an instrument with suitable sensitivity and specificity is non-trivial. One promising means of biomass, substrate (glucose, glutamine, glutamate), by-product (lactate, ammonia) and end-product monitoring uses spectroscopic sensors based on near infrared (NIR), mid infrared (MIR) or Raman spectroscopy.. These optical methods have many desirable attributes for bioprocess monitoring:

  • they are non-invasive and non-destructive,
  • do not consume the analyte or require sampling,
  • are capable of monitoring several analytes simultaneously,
  • provide continuous real-time measurements,
  • no additional reagents are required and
  • they do not interfere with cellular metabolism or the bioreactor environment.
Spectroscopic techniques can also monitor multiple components simultaneously, eliminating the need for multiple sensors (5). Modelling: Capturing Process Knowledge from Process Data With the ever-increasing volumes of process data acquired through new sensor technologies, data analysis and process modelling is now a pivotal element in ensuring that the effort and costs associated with deploying a PAT strategy are beneficial to delivering a high quality product in an economical manner. Collecting the data is the first step. It must be translated into process information and further transformed into knowledge of the process and design space in order to be of value and repay the investment. Based on the process knowledge garnered, the CPPs for the process are identified and mechanistic process models which describe their trajectories and behaviours can be developed. The objective of developing a model is three-fold:
  • the development of the model further drives the expansion of the underlying process knowledge,
  •  the final model can then be used as a virtual environment in which to explore the operating space without incurring a large experimental workload, and
  • the model can finally be used as an element of an advanced control algorithm.
Mechanistic models seek to describe phenomena in terms of the underlying fundamental principles, such as mass balances or reaction pathways. One of the most common mechanistic models used for bioprocesses is based on the Monod kinetic model (6). One of the major benefits of developing a mechanistic model over a data-driven model is that it may be used for predictive purposes because of the knowledge that it contains. It could, for example, be used to explore the effect of different feeding regimes or process set-points in an animal cell culture process without the need to run many iterative experiments. In effect, the experiments could be run in the virtual environment. Control: Harnessing Process Knowledge [caption id="attachment_16010" align="alignright" width="572"]Figure 4: Generalised process control model using a PAT approach intended to reduce process variability Figure 4: Generalised process control model using a PAT approach intended to reduce process variability (click to enlarge)[/caption] Biological systems have inherently large variations between batches. Controlling the bioreactor helps to reduce the variability and to protect product quality (Figure 4). However, bioprocesses are dynamically complex; therefore, their control is a challenging and delicate task. Process control attempts to influence the sophisticated metabolic reactions inside the cell by controlling the extracellular environment. The limitations associated with the standard process control solutions have led to significant interest in advanced process control (APC) across numerous industries. Model predictive control (MPC) is currently the most widely used of all advanced control methodologies for industrial applications. The underlying control principle is to apply knowledge gained from higher process understanding to mechanistically link critical product attributes with measurements of process variables supporting the institution of optimal process control. The essence of MPC is to optimise forecasts of process behaviour. This forecasting is accomplished with a process model, and, therefore, the model is an essential element of an MPC controller. Case Study: Feed Profile Optimisation of a Mammalian Cell Process The objective was to improve the process performance, by optimising the bioreactor feed profile by moving from a bolus to a continuous feeding strategy in order to prevent nutrient depletion and deliver a stable macro-environment for the cells. The results of this study are detailed in Craven et al., 2014 (7). Phase 1: PAT Monitoring Typically, the initial phase involves advanced monitoring of the current process with appropriate PAT technologies and offline sample analysis in order to maximise the process data available. The data analysis is used in order to understand the interdependencies between the parameters of interest and to identify supplemental experiments that may be required to further interrogate them. In this instance, Raman spectroscopy was used for the simultaneous, in-situ, real-time estimation of important process parameters including: total cell density (TCD), viable cell density (VCD), glucose, glutamine, glutamate, lactate and ammonia. A Kaiser RXN2 spectroscopic system was used. Chemometric partial least squared (PLS) calibration models were developed using the Umetrics Simca multivariate data analysis package. These models were used to translate the Raman spectra into useful information for the standard process. These data streams were then analysed in conjunction with additional offline data in order to understand the process better and to identify a direction for the process development. Some of the conclusions drawn were:
  • the cell line had a high specific glucose consumption rate,
  • the specific glucose consumption rate was proportional to the glucose concentration,
  • lactate production was proportional to glucose consumption, and
  • there was a fixed ratio between the glucose and glutamine consumption rates.
Therefore, glucose set-point control was identified as potentially beneficial as continuous delivery of the glucose feed would prevent glucose depletion and reduce lactate production by controlling the glucose concentration. Glutamine concentration could also be indirectly controlled, giving the additional advantage of reducing ammonia production due to reduced glutamine consumption and degradation. With glucose set-point control in mind, Phase 2 of the PAT-enabled process development strategy was progressed. Phase 2: Bioprocess Modelling In Phase 2, based on the process knowledge garnered, the CPPs for the process are identified and process models which describe their trajectories and behaviours are developed. In essence, the model is a mathematical relationship between the inputs, the initial values of the identified CPPs, and the desired outputs, CPP trajectories and/or CQAs. The exact nature of the model developed depends on the data available, the attributes of the system and the intended application. [caption id="attachment_16013" align="alignright" width="753"]Figure 5: Output of a model developed for a Chinese hamster ovary cell fed-batch process compared to offline samples and on-line Raman determined trajectories (click to enlarge) Figure 5: Output of a model developed for a Chinese hamster ovary (CHO) cell fed-batch process compared to offline samples and on-line Raman determined trajectories (click to enlarge)[/caption] Figure 5 illustrates the output of the process model and online Raman determined parameters for the standard bolus CHO fed-batch process. The model was developed to ultimately be a principal component of a control algorithm for the closed loop feed back control of glucose concentration in the reactor. Therefore, the emphasis was on describing the glucose requirements of the system. In this case, the model was a system of ordinary differential equations based on mass balances and rate equations to describe cell growth and the consumption or production of particular substrates and metabolites which influenced the overall glucose demand. Prior to implementing the process model as part of the glucose set-point control strategy, a series of process simulations were run. The purpose of the process simulations was two-fold. Most importantly, the simulation results gave direction to the experimental design during process development by providing a virtual environment in which to test and optimise a variety of operating conditions. For example, different feed compositions were explored to determine the optimum ratios and initial concentrations of glucose and glutamine in the feed medium and to calculate the feed profile required to deliver constant glucose concentrations (Figure 6). Additionally, the virtual environment was used for the verification of the model structure and performance by comparing the output against the existing as well as unseen experimental data sets. [caption id="attachment_16015" align="alignright" width="761"]Figure 6 Figure 6: Simulation results for glucose set-point control of a CHO cell fed-batch process. (A) The glucose concentration profile when (B) various feed trajectories designed to control glucose concentrations to a fixed set-point are simulated. (C) The resultant cumulative feed volume added over the course of the process. (click to enlarge)[/caption] The virtual environment determined the glucose feed profile required and automatically took into account the changes in specific glucose consumption rate associated with the lag, growth, stationary and decline phases of a bioprocess. Due to the non-linear nature of bioprocesses, transitions in rate for the feed profile required to deliver fixed glucose concentration is potentially non-intuitive and would require a number of experimental iterations. The virtual process environment significantly reduces this workload. Phase 3: Closing the Loop [caption id="attachment_16016" align="alignright" width="751"]Figure 7 Figure 7: Comparison of the simulated and experimental (A) glucose profiles, (B) federate profiles and (C) cumulative feed volumes for a CHO fed-batch process. The experimental feed rates were determined by MPC with a 12 h interval (click to enlarge)[/caption] In this case study, the control target was glucose concentration. The online process value (PV) was the on-line Kaiser Raman-determined glucose concentration. Although it was possible to measure the glucose reading every six minutes, a measurement interval of twelve hours was used in order to demonstrate the MPC functionality and to challenge the model. The greater the measurement interval the more accurate the process model predictions need to be. The in-house developed MPC uses the process model to calculate the system’s demand for glucose, via the embedded process kinetics and engineering principles, between measurement intervals and uses comparison of the PV to the set-point to correct for any drift or error that accrued over the preceding control interval. Overall, the glucose set-point control resulted in an increase in peak viable cell density (VCD) and the integral of the viable cell density (IVC) which is directly related to increased titre (Figure 8). CONCLUSIONS [caption id="attachment_16018" align="alignright" width="739"]Figure 8 Figure 8: Comparison of the (A) bolus and (B) continuous glucose concentration profiles and (C&D) resultant cell density profiles (click to enlarge)[/caption] The principal goal of process development and control under this new paradigm is to provide an efficient and well-understood process in which all sources of variation are defined and the critical ones are controlled. Specifically, product quality attributes should be accurately and reliably predicted over an applied process control space. This implies that CPPs are controlled to their respective target levels. Ultimately, the outcome is greater process understanding and the streamlined optimisation of the bioprocess, delivering improvements in titre, product quality and process reproducibility. Stephen CravenDr Stephen Craven earned his PhD from the School of Chemical and Bioprocess Engineering, University College Dublin, where he developed bioprocess models for mammalian cell fermentations and also developed and applied advanced control strategies to PAT enabled bioprocesses. Dr Craven currently works as a research engineer with the Applied Process Company (APC) Ltd, Dublin. APC is the world's leading pharmaceutical process engineering research and development company. References [1] BioPlan Associates Inc. (2014). 11th Annual Report and Survey of Biopharmaceutical Manufacturing Capacity and Production. ISBN 978-1-934106-24-2. [2] IMARC (2004). Global Biopharmaceutical Market Report (2010-2015), IMARC Report. [3] Enright S. andDaltonM. (2013). The Impact of the Patent Cliff on Pharma-Chem Output inIreland. Department of Finance. Working Paper No.1. [4] FDA. (2004). Guidance for Industry Guidance for Industry PAT — A Framework for Innovative Pharmaceutical. [5] Whelan J., Craven S., Glennon B., (2012). In situ Raman spectroscopy for simultaneous monitoring of multiple process parameters in mammalian cell culture bioreactors. Biotech. Prog. 28(5): 1355-1362. [6] Craven S., Shirsat N., Whelan J., Glennon B. (2012). Process model comparison and transferability across bioreactor scales and modes of operation for a mammalian cell bioprocess.Biotechnol. Prog. 29(1): 186-196. [7] Craven S., Whelan J., Glennon B. (2014). Glucose concentration control of a fed-batch mammalian cell bioprocess using a nonlinear model predictive controller.J. Process Control 24(4), 344.