Bioprocess and Fermentation Monitoring (PDF Download Available)

PART V PROCESS ANALYTICAL TECHNOLOGIES (PAT)

1469

69 BIOPROCESS AND FERMENTATION MONITORING Michael Pohlscheidt, Salim Charaniya, and Christopher Bork Genentech Inc., Manufacturing Science and Technology, Oceanside, California, USA

Marco Jenzsch and Tim L. Noetzel Roche Diagnostics GmbH, Pharma Biotech, Penzberg, Germany

Andreas Luebbert Martin Luther University, Halle, Germany

69.1

INTRODUCTION

The monitoring and control of processes is of key importance in all industries. Effective methods of monitoring are required to develop, optimize, and maintain processes at a maximum efficiency and desired product quality. Biotechnological processes are no exception to this (1,2). Biotechnology processes are used to produce a large variety of products, such as primary and secondary metabolites, cells, tissues, vaccines, and therapeutic proteins (2). Different host cell systems are used in the modern biotechnology, for example, bacterial cells, plant cells, and eukaryotic cells, with specific requirements for bioreactor design, media composition, and process control. Especially the production of recombinant proteins and antibodies has become a major source of revenue during the past 30 years, which are typically produced by genetically engineered mammalian cells. The cultivation of mammalian cells requires, among other factors, complex media composition, specialized bioreactor design, and the control of various parameters in narrow ranges to obtain the desired productivity and product quality. Therefore, more research has been devoted to develop specifically designed sensors, sampling strategies, and integrated data management systems to allow better and more detailed process monitoring (3).

Advanced analytical technologies are required in all phases of a product and process life cycle. During the development stage, it is needed to create knowledge and understanding of the process and therefore optimize the process with accelerated timelines (e.g. high throughput screening and design of experiments). During process characterization and validation, it is essential to understand and determine control limits, set points, and critical process parameters, which have an influence on process performance and, more importantly, affect product quality. In routine production, it is essential to control the process at the set points within the operating ranges. These process parameters are defined during process development and characterization to achieve the desired product quantity and quality and support ongoing process monitoring/validation as required by the health authorities. The process analytical technology (PAT) framework published by the Food and Drug Administration (FDA) in 2004 describes PAT as a mechanism to design, analyze, and control pharmaceutical manufacturing processes through the measurement of (critical) process parameters that affect quality attributes of the product (4). The concept is based on building quality through deep understanding and control of the process, especially critical quality attributes and their interactions/dependencies to other parameters. This requires a comprehensive detection system to measure, analyze,

Upstream Industrial Biotechnology: Equipment, Process Design, Sensing, Control, and cGMP Operations, Volume 2, First Edition. Edited by Michael C. Flickinger. © 2013 John Wiley & Sons, Inc. Published 2013 by John Wiley & Sons, Inc.

1471

1472

BIOPROCESS AND FERMENTATION MONITORING

monitor, and control all important attributes of a bioprocess and the employment of feedback-based control systems, including integrated data management and analysis. Over the past several years, there has been a lot of effort in analytical technologies to enable these stringent controls (5). The chapter describes general operational aspects of sensors, sampling technologies, and methods of process monitoring, examples for different controls and advanced applications such as soft sensors and metabolic controls. Some aspects are based on a previous article published in 2010 (1). 69.2 GENERAL ASPECTS OF SENSORS AND MONITORING Several bioprocess sensors have been designed and are employed for routine applications. For bioprocesses monitoring, the following areas for physical, chemical, and biological parameters are commonly monitored (Fig. 69.1). For any measuring technique, there are a number of criteria that the measuring device must satisfy to enable its deployment in commercial bioprocess manufacturing. The following list of characteristic is not exclusive, but does cover the main aspects of sensor attributes and characteristics (see also Ref. 1 for more details). 69.2.1

Measurement Frequency and Costs

In most cases—especially in context of the PAT initiative—real-time measurements at a high frequency are highly desirable. However, in some cases, this is not necessary and the kinetics of the reaction, consumption, and evolution must be considered in determining the frequency of measurements. In addition, differences in the life cycle of the process and cost aspects must be evaluated during the selection of one or more monitoring techniques. For research and development (R&D) purposes, in general, more data are obtained and collected to gain deeper process understanding and optimize process robustness and yield. At this stage, higher costs are often more Osmolality Redox

Metabolites/substrates

Viscosity

pH O2

Volume/weight

Biomass/viability Bioprocess monitoring

CO2 Flow

Product concentration Product characteristics Impurities

Pressure

Genetic/metabolic analysis

acceptable compared with commercial bioprocessing. For routine operations, critical and important variables must be monitored and controlled within predefined intervals. A secondary aspect is cost; measurements that are resource and time intensive, but not critical to the process from a product quality and quantity perspective, may not be implemented. In general, high frequency at low costs, time, and resource consumption is desired, including automated data handling and analysis. For biological processes, the growth rates, substrate depletion rates, and evolution rates of products, by-products, and impurities are most relevant for the measurement frequency. Animal cells, for example, display a doubling time of approximately 1/ day. Bacterial cultures such as Escherichia coli display a doubling time of 20–30 min, also reflecting their metabolic activities. It becomes obvious that the measurement frequency to control feed rates and other physical parameters has to be much higher in bacterial cultures compared with animal cell cultures. Therefore, some fully automated systems have been developed for off-line or online analysis. In situ probes (discussed later) with fully automated analysis, such as spectral analysis, in situ microscopes, and off-gas analysis, have also been established. 69.2.2

Ease of Validation and Implementation

If operated in a regulated environment, the method and sensor should be easy to validate and qualify. Efforts required to validate and implement the sensor should not be complex as this could also affect project timelines, for example, implementation of a new sensor during a technology transfer project. Standard platforms that have been assessed toward this factor are often used for bioprocess operations at a large scale. In recent years, the aspects of leachables and extractables have become more important in the industry and should be considered as well because these can affect cells or even product quality attributes. 69.2.3

Easy to Use and Maintain

A sensor should be easy to maintain and easy to use (e.g. calibration and exchange of membranes). This may not be a primary focus during research and development. However, for routine operations, this is an important factor in terms of reliable execution: manufacturing schedule impacts for planned and predictable performance of operations. In general, calibration of the sensor and other preventive maintenance activities should be part of routine operations with intervals: calibration ranges and other requirements specified in operating procedures.

Stirrer speed Temperature

Figure 69.1. Major areas for bioprocess and fermentation monitoring.

69.2.4

Cleaning Aspects and Sterility

When operated in an aseptic or sterile environment, the sensor must be amenable to clean in place (CIP) and steam

1473

METHODS OF MONITORING

in place (SIP) without nonpredictable effects to performance of the sensor. This can be a threat because aggressive chemicals as well as high temperatures may be used to obtain sufficient cleaning and sterility/aseptic conditions. In biotech processes, temperatures in the range of 121–130◦ C are common for sterilization, and treatment of surfaces with hot base or acid is common for cleaning. The general acceptance and impact on lifetimes and performance for various probes can be found in vendor recommendations. Nevertheless, the standard probes currently applied in most routine (GMP) production biotech processes, such as pH or dO2 , are able to cope with these conditions quite well. Sensor drifts affected by caustic solution or heat treatment my occur, but are usually not an issue because most often these probes are operated in a single-use mode or at least only for a few batches with appropriate calibration cycles that assure the desired sensor performance. Beside, the possibility of in situ sterilization probes and their fittings into the bioreactor are a critical point with respect to the sterile boundary. This is one of the key aspects, which has to be considered when introducing new probes into a (GMP) manufacturing environment. 69.2.5

Reliability, Accuracy, and Reproducibility

A sensor, especially when used online, has to provide reliable performance over a large span of time. Mammalian cell culture processes, for example, have operating times of up to several weeks (or months in perfusion systems) and therefore the probe/sensor has to deliver reliable performance to avoid failures and batch losses. The sensor should always provide sufficient accuracy in the desired control range. A low error between actual value and measured value is a prerequisite for adequate control. Measurement should be reproducible and drift only in an acceptable range for the measurement. 69.2.6

Linearity, Sensitivity, and Specificity

A measurement signal in most cases is directly proportional to the concentration of the component. Sometimes, there need to be corrections, and sample dilution has to be performed to overcome the effect of nonlinearity. A sensor

has to deliver appropriate sensitivity to measure changes in the concentrations. The lowest level of detection is related to the sensitivity and the signal-to-noise ratio. Specificity is a statistical measure of how well a test correctly identifies the negative cases or those cases that do not meet the condition under study. Specificity is defined as the number of true negatives divided by the total sum of true negatives and false positives. A sensor with high sensitivity and specificity is capable of identifying and possibly quantifying a small signal, while still differentiating between a true signal and noise. 69.2.7

Response Time

The measurement of any process variable will entail a time delay between change in the parameter and display of the measured value. The response time should be appropriate for the progress of the bioprocess, particularly if the measurements are linked to a control action. For tight control limits and fast reaction times, an off-line sensor may not be appropriate. In situ measurement or online monitoring might be more appropriate. The response time of the sensor also depends on the control requirements of the bioprocess. For example, in a high cell density culture, dissolved oxygen (DO) will be depleted in a few minutes. Hence, DO is almost always monitored using an in situ sensor with a response time in seconds. The importance of these aspects and factors varies and is different from R&D to a routine manufacturing application (6). Table 69.1 shows the importance of different aspects during the life cycle of a process (+ = low; ++ = medium; + + + = high).

69.3

METHODS OF MONITORING

In principle, analytics can be performed in two ways: withdraw a sample and perform off-line analysis or by integrating the sensor in the process, for example, bioreactor, peripheral equipment, and external loops, to enable online or in situ monitoring (Fig. 69.2). In some cases, a combination with auto sampling devices can lead to an “in situ sampling and ex situ sensor” (3,5).

TABLE 69.1. Aspects of Sensor and Their Importance During the Life Cycle of a Process from Research to Commercial Manufacturing (6) Factors/Aspect

Research

Process Development

Manufacturing

Reliability, accuracy, reproducibility Selectivity, sensitivity, linearity Calibration Ease of validation and implementation Cleaning aspects and sterility Analysis frequency and costs Robust and easy to maintain

+++ +++ +/++ ++ +/+ + + ++ +

+++ +++ ++ ++ +/+ + + ++ ++

+++ +++ +++ +++ +++ +++ +++

1474


69.3.3

In situ analysis

Online analysis

Off-line analysis Bioreactor

Figure 69.2. Different methods of monitoring and sensor interfaces with the process.

69.3.1

Off-Line Monitoring

For off-line monitoring, a sample is withdrawn from the bioreactor/vessel and analyzed after sample preparation in a suitable device. The preparation and handling of the sample is of crucial importance because it may affect the accuracy of the measurement. For example, for cell counts, homogeneous mixing after sampling and short lead times can be important. For pH and CO2 , rapid measurement is needed to avoid drifting due to off-gassing. This is especially important for pH to avoid offsets between in situ and off-line control measurements because pH is often controlled in a narrow range for bioprocesses. An important aspect (especially for bioburden monitoring) can be the aseptic treatment of the sample. For all off-line samples, detailed procedures, trained staff, and suitable laboratory are required. Off-line analytics always has a delay and affects the frequency of the measurement.

69.3.2

In situ or In-Line Monitoring

In Situ and in-line monitoring involves the use of sensors placed directly in the vessel or the flow lines associated with it or the unit operation (e.g. conductivity or pH monitoring in purification steps). The use of in situ sensors are well established, especially for chemical or physical parameters (pH, redox potential, O2 , CO2 , conductivity, and turbidity) (7). In Situ monitoring allows rapid measurement in high frequency and therefore enables real-time measurement and direct control. Robustness, long-term stability, aseptic design, and SIP/CIP capability are a prerequisite (and sometimes a pitfall) for these sensors and probes. Sometimes, in situ probes can be placed in external loops. The use of bypasses and additional peripheral equipment always adds potential risks to the process (e.g. sterile barriers, pumping of cells and protein and associated shear force, and cells outside a controlled environment) and therefore need to be carefully considered.

Online Monitoring

A compromise between in situ and off-line monitoring can be found with online monitoring, where a sample is automatically withdrawn and analyzed. One example of this has been the development (since the early 1980s) of flow injection analysis (FIA). This is a liquid handling technique that has proved flexible and adaptable to most chemical and biochemical reaction procedures (8), representing an effective compromise between the desirability of in situ monitoring and the technical ease of off-line measurements. The principal advantages of online methods are that sensor sterilization can be readily accomplished, sample pretreatment (e.g. gassing, dilution, and removal of interferences) is readily achievable, and sensor calibration can be built into the system. The main disadvantages are a need for an effective and reliable sampling system and the fact that the signal is discontinuous; the frequency of measurement is determined by the overall FIA design and the inherent limitation of the approach (3).

69.4 69.4.1

SENSOR, DEVICES, AND TECHNOLOGIES pH Measurement

pH is a critical factor in most bioprocesses (5,6). As the impact on performance and in some cases impact on product quality has been described, an in situ monitoring and control is needed. In addition, a risk mitigation strategy is widely used by comparing in situ measurement with periodic off-line measurements. The most common form of pH sensor used for fermentation monitoring is based on the electrode design introduced by Ingold in 1947. The detailed design of a pH electrode and functional principles can be found elsewhere (1). A large variety of pH probes based on the electrochemical design are commercially available. pH probes can be autoclaved and cleaned in place. However, a reduction in lifetime and robustness needs to be taken into account. Steam sterilization can affect the glass membrane potential, thereby adversely affecting the robustness of these conventional pH probes (9). Also, optical pH sensor, which uses pH indicators immobilized on waveguides such as optical fibers, is available. Measurement of pH changes can be detected with a good precision by the change in absorption or in fluorescence. With the development of disposable bioreactors, the need for disposable sensors has also evolved. Several companies have developed disposable sensors for pH, DO, and dissolved CO2 (DCO2 ). In these sensors, a fluorescence dye sensitive to the analyte is immobilized in a matrix or patch, which can be miniaturized to a few millimeters. This enables monitoring and control with presterilized/γ -irradiated systems, even of miniature system (e.g. shake flasks and microplates) (1,5,6)

SENSOR, DEVICES, AND TECHNOLOGIES

TABLE 69.2.

Types and Principals of Temperature Probes

Type

Principle

Thermistors

Resistance thermometer

Bimetallic thermometers

69.4.2

Type of resistor measuring the change in resistance. The output response from thermistors is of a nonlinear temperature versus resistance curve, with the resistance decreasing as the temperature increases Based on the changes in the electrical resistance of metallic conductors (mostly platinum), with changing temperature (10). A platinum wire of 100- resistance at 0◦ C is typically used—platinum sensors are stable under both sterilization and fermentation conditions Consist of a bimetallic helical coil surrounded by a protective tube or wall. The coil winds or unwinds with changes in temperature and causes movement of a fixed pointer (11)

Temperature

Temperature is a well understood and a well-controlled parameter in bioprocesses. Bioprocesses are monitored and controlled usually in a tight range of 0–121◦ C, including sterilization cycles. Several devices and principles are available to control temperature. Table 69.2 shows three of the most common applications. Other applications and more details can be found in Refs 1 and 11. 69.4.3

Pressure

Pressure is an important control parameter because it affects not only the bioprocess but also safety. Pressure, in general, influences the saturation concentration of gases dissolved in the liquid phase. Therefore, the conditions for calibration (if not automatically pressure corrected) need to be taken into account. In some bacterial fermentations, pressure increase is sometimes used to enable higher mass transfer. For cell culture applications, this can lead to higher carbon dioxide levels, which can negatively influence the process. Pressurizing the fermenter also mitigates the risk of undesired contaminations. Pressure monitoring is also essential for sterilization of bioreactors. Several devices and principles are described (e.g. Bourdon tube pressure gauge

1475

and strain gauge). Most common applications are piezoelectric manometers and diaphragm-type sensors. 69.4.4

Viscosity

Information on rheology, or viscosity, can help in ensuring the efficiency of a biological process and understanding the behavior of mixture, flow, mass transfer, and heat transfer. Animal cell culture media are mostly Newtonian fluids. However, for high protein concentrations in formulation processes and bacterial fermentations for the production of polysaccharides and DNA, the liquid can depict non-Newtonian characteristics. These can have severe inputs on power characteristic, mass transfer, shear stress, pumping, bioreactor design, and so on. The viscosity can be determined by using commercial available viscometers, for example, cone and plate viscometers, coaxial cylinders viscometers, and impeller viscometers. Viscosity is especially important in the production of high protein concentration, for example, in subcutaneous formulation of antibodies (12). 69.4.5

Redox Potential

Monitoring the redox potential of a bioprocess medium can provide information about the equilibrium between oxidizing and reducing species (electron acceptors and donors, respectively) present. The redox potential plays a significant role in understanding and optimizing secretion of proteins from mammalian cells (13,14). Typically, the metal electrode can be gold, iridium, or platinum; with platinum being the usual choice. The reference electrode is either Ag/AgCl or calomel. The redox sensor is linked to a pH meter that is fitted to provide readout in millivolts. During operations, the redox potential (measured between the metal and reference electrodes) varies as the logarithm of the ratio of oxidizable and reducible components in the media. The redox value varies linearly with the pH of the media and with the logarithm of DO tension via chemical reactions of the media components or culture components. This can lead to complication when trying to interpret the readout from the sensor. A full description of the use and problems associated with the redox sensor was outlined by Kjaergard (13). New and more reliable sensors have been published by Pluschkel and Flickinger (14), including a case study on mammalian cells. 69.4.6

Osmolality

Osmolality is a measure to describe the solute concentration defined as number of osmoles per liter of solution. The osmolality is therefore a result of dissolved components in a media. Freezing point depression, a phenomenon where the freezing point of a solvent decreases due to the presence of a non-volatile solute, is the commonly used method to

1476


determine the osmolality of culture broth in bioprocesses. Mammalian cells are typically cultured in conditions where the osmolality is in the range of 300–450 mOsm/kg. Higher osmolality can increase the specific productivity of recombinant proteins in mammalians (15). However, this is often accompanied by an impact on cell growth. Although not directly controlled, osmolality is an important parameter for monitoring mammalian bioprocesses. 69.4.7

Oxygen

Oxygen is an essential control parameter for all aerobic bioprocesses and can have influence on culture performance— the oxygen demand of cultures varies significantly. For cell culture, ranges from 20 to 200 mg/L/h have been reported (20). Impacts on metabolism and effects on cell culture have been described in high and low oxygen concentration—as well as on recovery process steps (21–24). The control of oxygen is a simple feedback loop. DO is measured by an in situ oxygen probe in most cases. Mammalian cell culture processes are typically robust to changes in DO in the range of 20–60% of air saturation. Transient low (60%) perturbations are also well tolerated by Chinese hamster ovary (CHO) cells, the commonly use host cell line for production of recombinant proteins. DO is typically controlled at 30–40% of air saturation to ensure sufficient “driving force” for oxygen transfer. The driving force is determined as the difference between saturation concentration (in equilibrium with the gas phase) and the control set point of DO (20). The control is based on the set point, and therefore, (dissolved) mass transfer for oxygen is controlled by gas flow rate, power input, and/or changes in gas composition. For cell cultures, the common practice is to keep gas flow constant at low power inputs and add more oxygen to the gas flow—this is feasible due to the low oxygen uptake rates (OURs) of mammalian cells. For bacterial fermentations, this is not the case. Here, high power inputs (>1 kW/m3 ) and high gas flow rates are used (combined with overpressure) to achieve high mass transfer characteristics in the bioreactor. Most differences for cell culture bioreactors and bacterial bioreactors are driven by this unit operation. Although high power inputs and high shear is desired in bacterial fermentations to achieve high mass transfer and low mixing times, this is the opposite in cell culture processes due to their shear sensitivity (lack of cell wall). In cell culture processes, power inputs between 10 and 100 W/m3 are typically used. In addition, several low shear–high mixing impeller types have been developed by retrofitting bacterial bioreactor operations. For high density cell cultures, often pure oxygen is used to achieve the necessary mass transfer rates without high power inputs. Specific sparger types (e.g. microsparger and open pipe sparger) are also used to avoid CO2 build up and high shear force due to bursting of gas bubbles at the air–liquid interface (16–19).

Oxygen (and carbon dioxide) can also be measured in the vent lines at the exit of bioreactors. This “off-gas” measurement can be used to decipher the metabolic state of the cell. The difference in the oxygen and carbon dioxide level between the inlet and the outlet of a bioreactor can be utilized to estimate the OUR and carbon dioxide production rate (CPR) in real time. These parameters, especially OUR, can be used to adapt feed rates (21,22). In general, off-gas analysis requires a certain change in gas concentration in order to estimate the consumption of oxygen and buildup of carbon dioxide. For bacterial cells the difference between inlet gas concentrations and off-gas concentration is big enough for differential measurement using mass spectrometers. For mammalian cells this is not the case. Therefore, off-gas analysis is not common in mammalian cell culture processes. Recently, due to availability of more sensitive detectors, examples of off-gas analysis have been reported for mammalian cells (21,22). Table 69.3 shows the most common methods to determine oxygen (dissolved as well as off-gas oxygen levels). A large variety of oxygen probes are available. The probes are easy to maintain and reliable, with sufficient specificity and reliability. The probes can be used in sterilization processes. Sometimes, long-term use after CIP/SIP can be an issue. Therefore, redundant electrodes are often used in processes to provide backup. However, preventive maintenance and preventive exchange of the in situ probes has been successfully applied of the past few years. For disposable application also, sensors are available based on fluorescence as described earlier (see pH). Beside, the critical role of oxygen in aerobic fermentation, impact of dissolve oxygen on product quality in downstream processing has also been reported. Pizarro et al. (23) describe the relevance of DO concentration during a refolding process of recombinant human vascular endothelial growth factor (rhVEGF). Kao et al . (24) described the reduction in interchain disulfide bonds in a therapeutic antibody, in a CHO-based manufacturing process after the harvest operations, resulting in observation of antibody fragments in protein A pool. As a result, certain measures were introduced, one of which was air sparging of the harvest tank (24). Despite their differences in operation and measurement principles, there are some constraints for electrodes. All electrodes need to have sufficient movement at the surface, including movement of oxygen through the bulk of the fermentation media and diffusion across the membrane and through the supporting electrolyte. An incorrect placing of the electrode in a highly viscous media displaying non-Newtonian characteristics might be an issue. If the electrode is situated in an area of quiescent fluid, the received signal may not be an accurate representation of oxygen partial pressure throughout the bioreactor. Pressure difference might play a significant role, as well


TABLE 69.3. Device and Principle for Oxygen Measurement Type

Principle

DO Probes

Optodes (DO)

Exit gas analysis

Based on galvanic or polarographic principle. Measurement of partial oxygen pressure (19,26–28) Optical sensors for oxygen are constructed using an immobilized fluorophore that undergoes dynamic quenching of the luminescence of a ruthenium complex. The change in fluorescence signal is measured by optical measurement device. Also available as single-use sensors There are a number of methods available for determining the oxygen concentration of exit gas from a bioreactor. Several of these are based on exploiting the strong affinity shown by oxygen to a magnetic field. Paramagnetic gases, such as oxygen, display a positive magnetic susceptibility. There are two types of detectors based on this phenomenon: the magnetopneumatic and thermomagnetic analyzers consist of a bimetallic helical coil surrounded by a protective tube or wall. The coil winds or unwinds with changes in temperature and causes movement of a fixed pointer (11)

as temperature control and compensation. Fouling of the sensor membrane surface by components of the bioprocess media can lead to erroneous signals. With continued use, this could include the growth of microorganisms on the membrane surface, especially when cultivating adherent cell lines. DO electrodes are also susceptible to signal drift during operation. In addition, gas bubbles and restricted flow toward the sensor can affect the received signal. Other important considerations are the effect of sterilization on the sensor and its calibration. Typically, calibration is carried out after sterilization, using medium gassing to obtain the full range of sensor values. However, in some cases, this might lead to excessive foaming or reduction/oxidation of media components and therefore needs to be considered carefully. For fast reactions, especially with high mass transfer and oxygen consumption rates, the response time of the probe and potential calculation need to be taken into account as well. 69.4.8

Carbon Dioxide

For many bioprocesses, the measurement of CO2 is an important feature. Media in bioprocesses is typically

1477

buffered by CO2 /bicarbonate buffer in addition to other buffers such as HEPES (16,20,21). Carbon dioxide is utilized for pH control in mammalian bioprocesses. In addition, carbon dioxide is also required for cellular processes such as fatty acid synthesis. At smaller scales, such as spinners and shake flasks, mammalian cells are typically cultured in incubators with 5% CO2 . However, increased levels of dissolved carbon dioxide can inhibit growth and reduce the production of secondary metabolites. Especially in cell culture processes, carbon dioxide is critical at certain concentration and has demonstrated impacts on product quantity and process performance. Hence, DCO2 is typically maintained between 5% and 10%. Cells produce CO2 during respiration, which is stripped off by air (or oxygen) sparged in the bioreactor. Carbon dioxide build up must be considered carefully during scale-up. Due to geometric similarity during scale-up, the surface-to-volume ratio decreases. Hence, the volumetric air/oxygen flow rate is reduced to avoid excessive foaming. For oxygen supply, this is addressed by supplying oxygen-enriched air or pure oxygen. However, this causes a reduction in CO2 stripping rate, resulting in an increase in DCO2 concentration. DCO2 levels above 20% are often growth inhibitory for mammalian cells. Hence, the aeration strategy should be carefully considered during scale-up of bioprocesses. In addition, the concentration of bicarbonate in the medium will affect the DCO2 levels. Carbon dioxide probes are commercially available. In principle, these are steamable and cleanable as described for pH and DO probes. These function in a similar way to oxygen probes and are based on a pH sensor immersed in a saturated bicarbonate buffer, separated from the bioreactor fluid by a hydrophobic membrane. DCO2 gas molecules diffuse from the bioreactor media, through the hydrophobic membrane into the carbonate buffer (e.g. InPro 5000i by Mettler Toledo). In situ fiber optic probes are also available for DCO2 measurement (e.g. YSI 8500 by YSI Life Sciences). 69.4.9

Weight and Liquid Levels

Weight and liquid levels are widely used to measure contents, nutrient additions and consumptions for mass balances, addition of base, flow rates, and so on. In most cases, load cells (measuring the strain on the device) or direct liquid level measurements are applied. The liquid level is also important from several perspectives (e.g. safety and total mass of product). In addition to load cells, several methods to determine liquid levels are known (e.g. measuring hydrostatic pressure) (29). If the culture media shows excessive foaming, the exhaust filters may foul. Hence, conductivity sensors are widely used to monitor foam formation. If liquid or foam levels rise and contact the nonisolated tip of the probe,

1478


the liquid or foam acts as an electrolyte—in most cases, this is used to control foam by addition of antifoam. Capacitance sensors detect the relative dielectric constant of the media compared with air. Acoustic sensors operate via transducers that generate and detect ultrasonic waves. Guided wave radar technologies are used as well. Guided wave radar is based on the principle of time domain reflectometry (TDR). TDR is a measurement technique used to determine the characteristics of electrical lines by observing reflected waveform pulses of electromagnetic energy that are transmitted down a probe and the pulse is reflected when it reaches a liquid surface. The distance to the reflecting surface is determined by the return time of the pulse to the source (30). Also, techniques of measuring temperature differences in the vessel can be applied (1). 69.4.10 Flow Measurement and Control of Liquids and Gas Flows need to be measured at the inlet and the outlet of a bioreactor, vessel, filter, or column. Flow control is important for process robustness and safety. In upstream processes, flow control is applied, for example, in fed-batch processes (potentially combined with weight measurements) or the control of airflow. In recovery operations, it is important to maintain flow rates in chromatography and filtration operations—especially viral validations and filter integrity might be affected because flow can be correlated to pressure and therefore in most cases is controlled. Calibrated rotameters or flow meters are typically used for flow measurement. Also, thermal mass flow controllers play an important role. For liquids, the characteristics of liquid (especially viscosity and conductivity) and the process requirements (sensitivity, accuracy, sterility, etc.) are used to determine the measurement device. 69.4.11

Biomass

In bioprocesses, the monitoring of biomass, cell densities (or cell volumes), is of crucial importance for the process. A wide array of techniques are available, usually carried out using microscopic method, volume measurements, dry weight, and so on (31,32). These are routinely performed off-line. In situ or online methods have been described as well (turbidity, in situ microscopes, capacity probes, etc.) for direct measurement of cell counts, and so on. Of particular interest is the development of real-time online methods. Generally, methods for determining biomass concentration can be divided into two classifications: direct and indirect. The former is based on determining the physical properties of the cell and its components. In contrast, indirect methods measure factors related to the cell and its activity (e.g. respiration, electrochemical behavior, and nutrient fluctuation).

Quantification of cell number concentration is performed ubiquitously in cell culture operations. Trypan blue exclusion method is most commonly used to differentiate between viable and dead cells, although other staining methods that use ethidium bromide or propidium iodide can also be used. Viable cell density (VCD) and viability (percentage of total cells that are viable) can be measured either manually (using hemacytometer) or using automated cell counters (e.g. Bioprofile® CDV from Nova Biomedical or the Cedex HiRes Cell Analyzer from Roche Innovatis), although the latter results in greater consistency. Automated cell counters can also provide statistics related to cell size and cellular aggregation levels. Cell concentration can also be measured indirectly by cell volume. A fixed volume of cell culture fluid is typically centrifuged in a graduated tube at a predetermined gravitational force for 5–10 min to estimate the volume of packed cells, referred to as the packed cell volume (PCV). In exponential growth phase, when cell viability is high, PCV is strongly correlated with VCD. However, due to the inability of PCV to differentiate between alive and dead cells, the correlation between PCV and VCD is weak at low cell viabilities. A primary disadvantage of these methods is their off-line nature. For mammalian cell culture process, a culture sample is drawn, typically every 12–24 h for each measurement. This limits the frequency at which cell growth can be monitored. Developing technologies to monitor cell growth in real-time has been an aggressive pursuit in the past two decades. Methods based on nuclear magnetic resonance spectroscopy, conductivity, capacitance, fluorescence, and optical absorbance have been tested in bioprocess (for review see Refs 5 and 32). Among these methods, capacitance-based cell density measurement appears promising. Cells behave as capacitors due to the presence of charged molecules both inside (cytoplasm) and outside (culture broth), separated by a plasma membrane. Capacitance, measured by application of an electric field, is directly proportional to the cell concentration. An important advantage of the method is its ability to differentiate between live and dead cells, because dead cells without intact membranes do not contribute to charge polarization. Good correlations between VCD and capacitance measurement have been demonstrated in several reports (reviewed in Ref. 33). Real-time viable cell mass estimation was used to optimize the perfusion rates in CHO-based recombinant protein production process (34). For large-scale applications, the sensitivity, especially at low cell densities, and linearity of capacitance-based cell density measurements should be carefully considered, because nonlinear correlations between capacitance and cell mass have also been reported in the fermentation of yeast (35). In-line optical density probes have also been employed to monitor total biomass concentration in


1479

TABLE 69.4. Examples for Optical Probes Used in Mammalian and Bacterial Fermentation Processes (5)

Aquasant AF44 SIR BE 2100 InPro 8100 Optek

Mechanism

Application

Back scattering Back scattering Back scattering Forward scattering

Bacterial and mammalian Bacterial Bacterial Not specified

bioprocesses. Table 69.4 shows the examples of optical probes in bacterial and mammalian fermentations. These probes rely on light scattering and absorbance properties of cells to estimate cell concentration. However, their application is limited due to their inability to distinguish between viable and dead cells. An in situ microscopic system was also developed for CHO-based process in a feasibility report (36). Despite its distinct advantage to enable real-time microscopic examination, a staining method could not be used with this technique to measure cell viability. Cell concentration can also be measured real time indirectly using indicators of cellular metabolism such as oxygen consumption rate, ATP production rate, or CO2 production rate (32). Among these indirect methods, oxygen-consumption-based estimate is most promising because DO probes are routinely used in cell culture unit operations. Empirical correlations of oxygen transfer coefficient (kL a) in stirred-tank bioreactors can be used to estimate approximate OUR. OUR can also be estimated periodically in the bioreactor by ceasing oxygen supply for short durations. Alternatively, the output of DO controller can be correlated to VCD (or PCV) in a bioreactor. As the cells divide and consume more oxygen, the output of the DO control will increase to maintain a constant DO level in the bioreactor, as shown in Fig. 69.3. However, DO-based methods assume that the specific oxygen update rate is relatively constant. Although this parameter depends on the physiological state of the cells, it is likely to be relatively constant during growth phases in seed and inoculum trains. Care should be taken, however, as these correlations may be cell line specific. Exit gas analysis can be applied as well, while using OURs displaying more accuracy. The “health” of a cell culture operation can be monitored by other indicators such as viability and apoptosis. Assays based on extracellular quantification of lactate dehydrogenase (LDH), the reversible enzyme that converts pyruvate to lactate, is a commonly used method to estimate viability indirectly. Viability is measured from the extracellular levels of LDH in membrane-compromised, damaged cells. However, this is a very indirect measure for viability. Programmed cell death or apoptosis can also be used to understand the physiological state of mammalian cells. High recombinant antibody production levels, often

1 Normalized output

Sensor

0.8 0.6 0.4 0.2 0 20

40

60

80

100

120

Time (h)

Figure 69.3. Robust correlation between DO controller output (black line) and PCV (open black square) for a CHO process. The Y -axis is normalized. DO output is represented as the ratio of the controller output (%) and the maximum output recorded at this bioreactor stage based on more than 100 batches. PCV is represented as percentage of the total volume.

accounting for more than 20% of cellular protein levels, can result in stress responses that induce apoptosis. A variety of colorimetric assay kits are available to detect early as well as late apoptotic cells. Many of these assays quantify the active levels of one of the many caspase enzymes that are activated during apoptosis. These assays can be used in conjunction with a flow cytometry device (such as a Guava Easycte) to observe distribution of cell populations. A flow injection instrument used in conjunction with flow cytometry allows real-time monitoring of cell subpopulations (38). Broger et al . (37) characterized different high and low EGF-producing populations of Pichia pastoris in small-scale bioreactors using real-time flow injection flow cytometry. Flow cytometry is a measuring technique based on the irradiation of a flowing sample solution (containing a cell population) with a suitable light source, followed by monitoring of the scattered or adsorbed light. In addition, fluorescence can be used as the measuring parameter. This technique can be used to ascertain a number of cellular features, such as the accumulation of cellular components (e.g. DNA, RNA, and proteins) and cell dynamics (e.g. cell size distribution). Furthermore, flow cytometry can be used to differentiate and quantify a range of species populations present in a mixed medium. Although such process

1480


instrumentation can be very valuable to further our understanding of cell population heterogeneity in bioprocesses and its impact on process variability, their use in large-scale manufacturing processes has been limited thus far. Their use is often limited to screening and selection of high protein producing cell lines during cell line development. 69.4.12 Monitoring of Other Process Indicators and Product Measuring and analyzing various metabolic parameters during fermentation is crucial for process control, to achieve and maintain a reproducible fermentation process, and to gain a deep process understanding to characterize, validate, and further improve fermentation processes. Conventionally, substrates, specific proteins, and the metabolites of fermentation media are analyzed using different methodologies depending on the nature of the analyte: that is, amperometric (enzyme membrane based), enzyme linked immunosorbent assay (ELISA), and/or high performance liquid chromatography (HPLC). For mammalian cell culture processes, the parameters typically monitored are glucose, lactate, ammonium, and glutamine. Glucose, lactate, glutamine, and glutamate are commonly measured using an enzyme-based biosensor that converts the substrate to hydrogen peroxide, which is oxidized to produce an amperometric signal proportional to substrate concentration. Examples include Nova Bioprofile Flex (Nova Biomedical) and YSI biochemistry analyzer (YSI Life Sciences). These bioprocess analyzers are most commonly used as off-line, standalone instruments for periodic nutrient measurements. However, autosampling technologies have been developed and successfully applied for cell culture. For bacterial fermentations, autosampling is challenging due to high viscosity, cleaning requirements, and so on. In all cases, comparable performance and reliability needs are to be demonstrated. The Nova Bioprofile Flex has been applied with autosampling technologies. It can combine three analyzers (metabolite—glucose, lactate, glutamine, glutamate), cell counter, as well as gases and electrolytes (pH, PO2 , pCO2 , ammonium, potassium, sodium) by using a reactor valve module. Details can be found in Refs 5 and 39. Methods based on photometric quantitation of nutrients such as glucose, glutamine, and other parameters have recently been commercialized (e.g. Cedex Bio by Roche Diagnostics GmbH Germany) to overcome potential disadvantages of other technologies. The Cedex Bio is shown in Fig. 69.4. In addition to the already mentioned substrates and metabolites, the Cedex Bio can analyze IgG concentration from cell culture broth without any small-scale purification. LDH activity can also be determined using Cedex Bio. LDH is a cytosolic protein released on the lysis of cells,

which can be representative to the amount of released proteases, glucosidase, nucleases, or host cell protein (HCP) detrimental to product. Other chemical parameters such as sodium and potassium concentration are also periodically monitored. As all these parameters are only measured once every 12–24 h, it is desirable to have a high degree of accuracy and precision in their measurement. However, the level of accuracy necessary for process control is different for different parameters. For example, if feed addition is based on controlling glucose at low concentrations, then the instrument must be capable of measuring glucose concentration in the low range (e.g. 0.1–2 g/L) accurately. On the other hand, parameters such as sodium and potassium concentration are typically not utilized for processing decisions, but are useful for troubleshooting purposes. A high degree of accuracy may not be necessary for such parameters. 69.4.13 In Situ Process Monitoring Using Spectrometry Autosampling can minimize disadvantages associated with off-line analysis and sampling procedures. However, this poses risks to sterility and robustness, while increasing complexity. In Situ sensors are therefore more advantageous. Several technologies are being developed for real-time estimation of nutrient levels. Technologies including Raman spectroscopy (41,42), infrared spectroscopy (43), UV spectroscopy (42), and fluorescence spectroscopy can monitor a wide array of species, including glucose, lactate, various amino acids, and NAD(P)H. Near infrared (NIR), mid infrared, and Raman spectroscopy have demonstrated successful applications in mammalian cell culture, yeast, and bacterial fermentations (5). In most cases, multivariate data analysis of large data sets from historical fermentations, model evaluation, and training were necessary to establish suitable models. Among all these technologies, NIR spectroscopy offers practical advantages with respect to cost and ease of use (for review see Ref. 44). However, interpretation of NIR absorbance spectra is rendered difficult due to the overlapping spectra of multiple species and the absorption of water in the NIR spectrum. Dimensionality reduction methods such as principal component analysis (PCA) and partial least squares (PLS) are commonly used to identify spectral correlations and express them as linear combinations of a small number of orthogonal dimensions. Pattern recognition methods [e.g. artificial neural network (ANN)] are also necessary to generate calibration or training data sets. 69.4.14

DNA Microarrays

The technology to examine the gene expression level of thousands of genes in a host cell (e.g. E. coli and CHO),


1481

Figure 69.4. Cedex Bio by Roche Diagnostics GmbH.

using a single chip called DNA microarray was developed less than two decades ago (45). A DNA microarray comprises thousands of probes, each of which is designed to bind with messenger RNA (mRNA) of a particular gene with high specificity. The probe is designed to be complementary to the nucleic acid sequence of the mRNA of a gene. DNA microarrays are commercially available for several species (e.g. mouse and human) from Roche Nimblegen, Affymetrix, or Agilent Technologies. These oligonucleotide microarrays typically comprise 25–60 base pair-long probes that are deposited on the microarray surface by in situ synthesis using photolithography Microarrays provide a valuable tool to decipher the physiological basis of traits that are biopharmaceutically relevant, such as high productivity of recombinant proteins in mammalian cells. Through direct comparison of the host cells under high productivity conditions and control conditions, the genes that are differentially up- or downregulated under the high productivity conditions can be identified. The biological function of the differentially expressed genes (e.g. role in cell cycle and protein synthesis and secretion) facilitates the correlation between the molecular snapshot obtained using microarray and the observed phenotype. Microarrays have been utilized as tools to identify the gene expression signature associated

with high recombinant protein producing cells lines (46), as well as high productivity conditions (47). These signatures provide useful targets for engineering cells with relevant traits. The applications of this technology in development of biopharmaceutical processes is reviewed elsewhere (48). Real-time quantitative polymerase chain reaction (Q-PCR) can also be used to quantify the transcript level of genes. Although Q-PCR assay has greater sensitivity and accuracy compared with DNA microarrays, it has significantly lower throughput (e.g. 96-well plate-based assay format). Q-PCR-based expression analysis is often useful when the number of genes are small and known (e.g. all the genes encoding proteins involved in a specific metabolic pathway). Microarrays can be used as diagnostic tools in cell culture processes. For example, investigations associated with lower productivity and adverse product quality trends during scale-up, technology transfer, and routine manufacturing can utilize a genome-scale survey tool such as microarray to facilitate product and process understanding, leading to possible corrective actions. Recently, the entire 2.45-Gb genome CHO-K1 parental cell line has been sequenced, which is publicly available (49,50). This development is likely to accelerate the use of whole-genome “omics” tools (e.g. a commercially

1482


available DNA microarray for CHO) in a wide array of bioprocess applications. Advent of massively parallel sequencing technologies has promoted development of several sequencing-based high throughput assays such as detection of noncoding microRNAs (79) and ChIP-Seq for detection of DNA sites that interact with proteins. As the cost of high throughput sequencing plummets, additional applications in cell engineering and bioprocessing will emerge. 69.4.15

Chromatography and Recovery Operations

After production in the fermentation process, the desired product has to be purified. The major goal of the recovery operations is to remove impurities (DNA, HCPs, etc.), potential viruses, and by-products. Therefore, chromatography has been widely used in the industry. The principle is based on the separation of a mixture of compounds by specific interactions with the matrix of a mobile and stationary phase. Process monitoring for recovery operations is usually made up of two components—in-process analytical tests and real-time measurements via in-line sensors and probes. In-process analytical tests are designed to provide information on product quality attributes at various stages throughout the purification process. Some of these tests (i.e. pH or UV absorbance) can be performed quickly by the operators on the manufacturing floor with little set-up time, while others (i.e. bioburden) may take several days to obtain results and usually require the expertise of a separate analytical group. Currently, most real-time measurements are designed to provide information about the process. With modern manufacturing control systems and data historians, large amounts of continuous data can be collected during a purification process. Sensors located on the process equipment provide information about key process indicators such as pH, conductivity, pressure, and UV absorbance. These process indicators can be visualized during the process itself or examined retrospectively and provide information on the performance of chromatography steps or filtration operations. They can also be trended versus historical limits to identify subtle process changes over time. Although most real-time measurements are currently being used to provide data about the process performance, recent advances in PAT can now allow a direct correlation to product quality as well. To control the process, several analysis and detection methods have been established. Especially the production of therapeutic proteins and antibodies requires a wide range of analytical methods to determine the concentration, purity, identity, integrity, and activity. A large number of operations and detectors have been developed for these purposes (flame, UV detectors, electrochemical detectors, IR and refractive index detectors, conductivity, and mass spectrometry). Also, a large number of chromatographic methods are

available for sample analysis. Most of the chromatographic analysis is carried out off-line. Table 69.5 shows exemplary analytics performed in a pharmaceutical antibody production process. Details can be found in the study by Flatman et al . (51). A state-of-the-art monoclonal antibody process can be found in the study by Shukla and Thömmes (52). Some real-time testing of product quantity can be performed online. Other operation monitoring parameters can be performed online as well. For example, one factor that can have a significant effect on chromatographic performance is the quality of the column packing. The current standard procedure for testing the quality of a packed bed liquid chromatography column is to use a nonabsorbed tracer to perform a pulse-injection experiment. The injected tracer solution is assumed to be a Dirac pulse. The pulse exits the column as a peak due to axial dispersion. Plate number, N, describes the degree of the dispersion, which is influenced by the packing quality of the column bed. A related term, Height Equivalent to a Theoretical Plate (HETP), provides a measure of peak broadening in relation to the distance the tracer has traveled in the chromatography column. The mathematical definitions of N and HETP are given by the following equations: N = Vr2 /σ 2 HETP = L/N where Vr is the retention volume, which is defined as the volume that has passed through the column from the time when half the tracer is applied to the time when half the tracer has exited the column. In other words, Vr is the mean exit volume of the injected tracer, σ 2 the variance of the exit volume distribution, and L the column length. Based on the normal density function, the width of a curve at half peak height, Wh , is equal to 2σ (2ln2)1/2 . Because the peak generated by the tracer as it exits the column is assumed to follow a Gaussian distribution, N is usually calculated with the simplified formula shown in Fig. 69.5. In recent years, efforts have been made to directly use process chromatography data to determine column efficiency to achieve real-time monitoring (53,54). The common approach taken in these studies is to utilize information from step transitions between buffers of different conductivities to describe the same dispersion parameters as the traditional pulse-injection method. One method is to transform a breakthrough curve or a washout curve into a peak by taking the first derivative (53). The dispersion parameters are then derived from the peak position and shape. To avoid the inaccuracy of calculation caused by assuming a normal distribution, algebraic functions other than the normal probability density function were evaluated, and a function that can describe


1483

TABLE 69.5. Examples of Analytical Methods Performed in a Pharmaceutical Antibody Production Process (51) Product Characteristic

Analysis Property

Physical and chemical characteristics

Purity Integrity/molecular weight

Identity Potency/activity

Antigen binding Biological methods

Product-related impurities

Aggregation/fragments

Process-related impurities and contaminants

Host Cell Protein (HCP) Host cell DNA

Protein A Cell culture medium proteins Viruses Microorganisms Others: column, vessel, filters, bags, leachables/ extractables, cell, culture medium additives, reagents, chemicals, etc.

1

Response

N = 5.54 (Vr /wh)2 HETP = L/N

0.5

wh

0 Vr

Figure 69.5. Plate calculation via traditional pulse-injection method.

a large number of step transitions has been identified. However, arbitrarily assigning a predetermined function to represent unknown distributions has disadvantages.

Method Electrophoresis Reverse-phase HPLC Size-exclusion HPLC Electrophoresis Mass spectrometry Size-exclusion HPLC Light scatter Isoelectric focusing Peptide mapping Ion-exchange HPLC Immunoassay Cell proliferation Complement-mediated cytotoxicity Reporter-gene assays Electrophoresis Size-exclusion HPLC Immunoassay DNA hybridization Q-PCR DNA binding threshold Fluorescent—picogreen Immunoassay Immunoassay Q-PCR Electron microscopy In vivo/in vitro assays Bioburden Endotoxin—LAL test Various, for example, reverse, phase HPLC, ion chromatography, GC-MS, and spectral analysis

Depending on the column packing quality and the running conditions of the chromatography, there are transitions that differ significantly from the chosen function and cannot be adequately represented by it. In these cases, the forced fitting of transitional data to the function would cause loss of information. Another method is to treat the exit volume of the solution that is replacing the original solution in the column as a discrete random variable (54). The incremental change in a response signal, such as conductivity, serves as the frequency of each exit volume. The starting point for the transition occurs when 0 L of the displacing buffer has flowed onto the column. The conductivity recorded at this point corresponds to that of the buffer on the column at that time (Cmin ). After a sufficient amount of the displacing buffer has flowed through the column, the conductivity will reach a new equilibrium (Cmax ).Cmax is equal to the sum of Cmin and the definite integral of dC, which is integrated from C = Cmin to C = Cmax . To simplify the calculations, C

1484


Normalized conductivity

1 VR ≈

s ≈ 2

n

(Vi · ΔCi) i =1

HETP (cm)

3.8 B A C

3.2

n (Vi -Vr) ΔCi i =1 2

N = V r 2 / s2 HETP = L /N

D

UCL

E

LCL

40

120

200 280 360 Number of cycles

440

520

Figure 69.7. Control chart of protein A column HETP derived from transition analysis. UCL, upper control limit; LCL, lower control limit.

0 Volume (VR)

69.4.16

Data Technologies and Artificial Intelligence

Figure 69.6. Plate calculation via transition analysis method.

may be first normalized using the following equation: Cnormal =

(Ci − Cmin ) (Cmax − Cmin )

As shown in Fig. 69.6, Vr and σ 2 can be calculated from the transition curve using the rectangular approximations of their integral forms. As the numeric data used in the analysis is discrete, Vr and σ 2 can be calculated using the following equations: Vr = (Vi Ci ) σ 2 = (Vi − Vr )2 Ci Note that the influence of each Vi on the output values of Vr and σ 2 is weighed by the magnitude of the corresponding Ci . The largest value for Ci usually occurs when its corresponding Vi is close to Vr . The values for N and HETP are calculated using equations shown later. Both breakthrough curves and washout curves can be analyzed in a similar fashion. Monitoring the integrity of large-scale packed bed liquid chromatography columns using transition analysis can provide useful information about the process. Figure 69.7 provides an example of how the HETP value derived from the transition analysis of a protein A chromatography column changed over multiple cycles of processing. The values increased with time after initial column packing (A). Increased measurement variability was also observed as integrity decreased. This was later determined to be caused by the formation of a headspace on the column. When the top flow adapter was lowered to eliminate the headspace, the HETP values were restored to their original values (B). However, subsequent repacking of the column once again revealed rapid degradation of the column integrity due to insufficient consolidation of the resin during the packing procedure (C and E). The flow adapter was lowered again after the second packing (D) to improve column performance.

The data obtained have to be analyzed to derive powerful and scientific sound conclusions, define interdependencies, and derive enhanced control strategies. Several approaches have been discussed in the literature, for example, hybrid modeling (55,56) and multivariate modeling, including principle component analysis, PLS, and ANNs (57,58). Recent developments in the field of artificial intelligence have led to investigations into the use of such systems for improving bioprocess control, based on the received measurement output signals. This has included the use of both knowledge-based expert systems and neural networks, during bioprocess operation. Robotics is also an important way to manage millions of chemical assays. Undoubtedly, the adaptation of such “intelligent” systems will develop over the coming years and will play an important role in the precise control of bioprocess applications. Multivariate tools are widely used to increase process understanding and optimization. Application in fermentations (including raw material screening) to analyze metabolic effects, estimations of final productivity in bioreactors, root cause analysis tools, real-time process monitoring, and so on have been described. Also, in recovery processes, this technology has been reported to add value. Chromatographic operations are often characterized by different phases that produce specific patterns and signals. For robust automation, these patterns have to be detected and discrete decisions need to be made at key points. Structured models are frequently used to perform online diagnosis and fault detection, and chromatogram overlays are often used to monitor interbatch performance. Early detection of process perturbations makes it easier to identify root causes and adjust control parameters or correct deficiencies to ensure high product quality and prevent batch loss. One method that has been described frequently in the literature for analyzing production data and gaining process understanding is multivariate statistical process control (MSPC) (59–62). Several papers have described the use of MSPC for monitoring batch processes to not only ensure acceptable product quality but also improve overall process performance and output


(63–65). In MSPC, multivariate models are first developed for “good batches” using PCA and PLS. These models can then be used to evaluate the performance of new batches in real time using commercially available software packages. When applying MSPC to chromatographic process data, a large number of factors can affect when certain events occur. Variations in the mobile phase composition, gradient reproducibility, temperature variations, and column packing variability can lead to shifts in the chromatographic patterns making objective analysis and comparison difficult. The unsynchronized nature of chromatographic data, therefore, implies the need for chromatogram alignment before analysis. Fortunately, most alignment problems can be resolved by normalizing the data based on the cumulative volume of buffer passed through the column. More sophisticated alignment techniques, such as dynamic time warping and correlation optimized warping, have also been used (66,67). 69.4.17

Softsensors

From a process engineering perspective, a software sensor, or in short, a softsensor is an estimation algorithm for a quantity that cannot be easily detected online. Information from other online sensors being utilized in the process can be used to estimate this quantity. Hence, a softsensor associates sensor hardware and an estimation method, that is, a software routine (70). Typically, softsensors were developed for online measurements of biomass concentrations, product concentrations, or specific biomass growth rates. The topic has also been reviewed elsewhere (75). At production scale bioreactors, the options for online measurement devices are rather restricted. Only a few basic devices can be validated such as sensors for pressure, temperature, pH, DO, and the off-gas volume ratios of O2 and CO2 . Hence, softsensors are essentially a software challenge. Softsensors most often do not take their information from a single sensor. Instead, they pool and condense the information from several ones, thus performing a multivariate analysis of the bioprocess to provide accurate estimates. This gives them a decisive advantage over single sensors: By means of simple balances, data consistency checks can be performed. This is important because decisions can only be made on measurement data that have good accuracy. Thus, it is not only the possibility to measure quantities that cannot be measured directly that makes softsensors attractive, the software can also be employed to perform tasks beyond the actual estimation. For instance, the software can perform fault analysis with respect to the sensor and the process. The complexity of signal processing may vary from simple code conversion to sophisticated extraction of pattern hidden in the data. The focus in this section is the estimation part of softsensor. In this respect, the softsensors are classified by the sophistication of the model relating the

1485

available sensor signals to the quantities to be estimated. Two basic categories are distinguished: data-driven and first principle models. 69.4.18 69.4.18.1

Basic Estimation Models Data-Driven Models

69.4.18.1.1 Utilization of Correlations. Simple engineering correlations are traditionally the first approach when process quantities must be related to each other. For instance, to track the biomass within a bioreactor with quantities that are measured online, a first choice is to look for the respiration rate of the cells. The OUR or the CPR can reliably be determined from the volume fractions of O2 and CO2 in the reactor’s vent line. They are related to biomass x by a simple relation similar to Luedeking and Piret’s (77) expression: OUR = Yox μX + m0 X This relation can be used to form a softsensor by converting it into an ordinary differential equation in biomass and solving it using a simple Euler algorithm. With an initial biomass x0 = x(t0 ), the final expression to be iteratively determined at each sampling time ti for OUR and culture mass W measurements: OUR(ti )W (ti ) − m0 x(ti−1 ) x(ti ) = x(ti−1 ) + dt Yox An equivalent sensor can also be obtained for the relation between biomass and the CPR, which can be measured online as well. Computationally, with softsensors we are faced with two different problems, we must distinguish between the computing power required during the online estimations (measurements). For the above example, this is the computing power needed for the stepwise solution of the Euler formula. On the other hand, we must consider the implementation expenses. In this particular case, it is the identification of the parameter values Yox and m0 from a set of training data measured beforehand for the process under consideration by means of nonlinear optimization procedures. While the first task is usually time critical, the second is performed off-line using optimization routines and is thus uncritical concerning computing time. 69.4.18.1.2 Polynomials. An improvement can be obtained by pooling the information from several measurements. In the simplest case, the following linear relation can be used: X = a0 + a1 OUR + a2 CPR + · · ·

1486


The parameters can simply be determined with multivariate regression algorithms available in all standard software libraries. Simple improvements of this basic linear approach can be obtained when OUR and CPR are replaced by their cumulative rates cOUR and cCPR. This measure acts as a powerful noise filter. In addition, another cumulative signal that is usually available at fermenters can be added. This is the total mass of base added to the culture during pH control as measured online with a simple balance under the base reservoir. This enlarges the linear equation to x = a0 + a1 cOUR + a2 cCPR + a3 Base The parameters ai can be determined in the same way as before using a set of training data. As the cultivation processes are intrinsically nonlinear, the linear relations can be extended to polynomials. A first approach is x = a0 + a1 CPR + a2 OUR + a3 Base 2 2 2 + a4 CPR + a5 OUR + a6 Base CPR OUR + a8 CPR Base + a7 + a9 OUR Base + a10 CPR Our Base With respect to the parameters ai , this equation is still linear; hence, the parameters can once again be identified with multiple linear regression routines. 69.4.18.1.3 Principal Components. Another linear technique often applied in multivariate data analysis is PCA. This technique is particularly important when many different measurements are to be exploited within the analysis, because the main idea is to transform the input data matrix composed of the signals from n input variables into a matrix of signals for a lower number of latent variables (< n) that essentially carry the same measurement information (62,72,74). As with multiple linear regression, PCA can easily be performed using general PC software. Ready to used subroutines are included in all common software packages (e.g. MATLAB® subroutine princomp). As many online process variables in a fermentation process are correlated (e.g. lactate concentration and base addition), a notable reduction in variables is possible. However, this can only be performed at the cost that the latent variables or scores obtained no longer depict a physical meaning. In fermentation processes where the number of online variables is rather low, the benefit of such an approach is thus questionable.

69.4.18.1.4 Artificial Neural Networks. A general approach of representing nonlinear relationships between several input and output variables is using ANNs. Even simple feed-forward networks with a single hidden layer proved to be able to represent the mappings that are of interest in softsensors. The term artificial neural network can be considered a catchword that is derived from the dataflow in a very simple algorithm. In the case of the simple feed-forward network with a single hidden layer with m nodes, it loses its myth when represented as a simple expression in term of a logistic function 1/[1 − exp(s)]. For the biomass x x(t) = wh

1 1 + exp[−wi s(t)]

This is a simple deterministic scalar product. The coefficients wh , referred to as weights, form the first (row) vector, which is multiplied with a second (column) vector composed by elements that are logistic functions with argument wi s, where wi is a weight matrix with m rows and n columns and s(t) the n-dimensional input (column) signal vector at time instant t. The coefficients of this representation, compiled in the vector wh and the matrix wi , can be determined in various ways. The easiest way is using the nonlinear optimization routines offered by the various software libraries (in MATLAB® the equation can be fitted to process data with lsqcurvefit). An important point to be considered is that it is not reasonable to feed all available online signals to the input of an ANN biomass estimator. It must be made sure that every input signal will more significantly contribute information to the biomass estimate than just putting noise into the analysis. This requires knowledge about the cultivation process (Fig. 69.8). More sophisticated versions of ANNs can be constructed by transferring the idea behind PCAs into the domain of nonlinear relationships. This is done in “autoassociative artificial neural networks” (aANNs). As with PCA, a set of k < n latent variables is produced. For this purpose, a regular ANN with n input nodes is used. Then, in addition, these k signals are mapped back into the original data space by means of two further feed-forward network layers. After determining all coefficients or weights, the entire aANN should be able to map all reasonable input vector combinations of the process under consideration onto themselves. In this way, the layer with the latent variables is directly related to the input variables, that is, a nonlinear version of the linear PCA is generated. Different estimators are often compared in terms of a performance measure, which is usually an average estimation error. Most often, the root-mean-square error (RMSE) is used for this purpose (73). A quantitative


50

50

Viable cell density (x105 cells/mL)

Run 001

Run 002

40

40

30

30

20

20

10

10

0

1487

0

2

4

6

10

8

12

0

0

2

4

6

8

10

12

4

6

8

10

12

60

45

Run 004

Run 003

40

50

35 40

30 25

30

20

20

15 10

10 5

0

2

4

6

10

8

Process time [d]

12

0

0

2

Process time [d]

Figure 69.8. Viable cell count off-line data (symbols) from four individual batches of a CHO recombinant protein production process together with the online estimates of an ANN (lines). Input variables are several signals from the DO controller of the bioreactor as well as base consumption and reactor weight.

TABLE 69.6. RMSE Values for the Different Biomass Estimation Methods as Described in Text Estimation Procedure Applied to VCD Data From a Production Cell Culture Process Biomass estimation by a feed-forward ANN Polynomial regression Multivariable nonlinear regression Luedeking/Piret-like relationship and OUR measurements

Biomass Estimation Error, RMSE (%) 3.24 4.76 4.12 4.41

comparison of the various approaches discussed here is shown in Table 69.6. All the estimators mentioned can be evaluated online, while the corresponding parameter optimization procedures can be rather time consuming. The ranking of the methods mentioned clearly shows that the best estimation quality can be obtained with simple feed-forward ANNs with a single hidden layer. This result proved to be valid in microbial and animal cell cultures at

various scales from the small liter scale-up to several cubic meters. 69.4.18.1.5 Support Vector Machines. Support vector machine (SVM) is an attractive alternative to ANNs. Compared with ANNs, SVMs find the global minimum, what cannot be guaranteed with neural networks. The basic idea behind SVM is to map the training data from the input space into a higher dimensional feature space. This transformation allows constructing a separating hyperplane with maximal margins in the feature space. Consequently, although we cannot determine linear function in the input space to decide what class a given data element belongs to, we can easily determine a hyperplane that can discriminate between two classes of data. The accuracy of the SVM regression is at least about the same as the result obtained from the ANNs discussed earlier. 69.4.18.1.6 Bayesian Approaches. As the performance of computers increased significantly over the years, more computationally demanding algorithms can be employed for

1488


analysis of online data. Thus, more practical applications of Bayes statistics (71) appear in the literature. Instead of looking at single parameter values within an estimation algorithm, it considers probability density distributions of the parameters to characterize the actual knowledge about their values. When the density function is tight, that is, the variance is small; we have a rather concrete knowledge about its value. Bayes’ law shows how the current prior knowledge about a parameter θ becomes sharper when we employ concrete measurement data from the process to obtain an improved a posteriori density function. Additional techniques such as clustering, decision trees, and k-nearest neighbors can also be employed for mining bioprocess data (reviewed in Ref. 80). 69.4.19

exploited in the so-called observers or filters. The extended Kalman filters (EKFs) are an advanced variant that has been applied to recombinant protein production processes (68,76,78). EKFs are one-step-ahead predictors: At each data sampling time, the EKF determines a new estimate as a weighted average from the previous estimate, the value predicted by the model and the measurements. When, for example, the measurement values are considered more accurate than the model predictions, they influence the estimate more than the values predicted by the model. A scheme for online estimating the biomass (X) and substrate (S) concentrations as well as the specific growth rates (μ) in a fed-batch process from OUR, CPR, and Base is depicted in Fig. 69.9. Basis is the first principle model based on the mass balances for X, S, and W , under the assumption that specific growth rate is to be kept constant, is shown on the left. Measurements H performed online are shown on the right. The connection between both is established by means of the measurement equation. The measurement equation transforms the actual state vector c into the measurement vector H . The EKF then provides a weight matrix Kg , referred to as the Kalman gain, which determines a correction c to state cpred predicted by the model from the difference between the actually measured values H and those (Hpred ) predicted from the model, to provide a new estimate cest . An example is shown in Fig. 69.10 for an E. coli culture producing a recombinant protein.

First Principle Models

69.4.19.1.7 Virtual Plant. When there is not enough data available from the process, the measurement information used to estimate the quantity under consideration can be supplemented by a priori knowledge about the process in form of first principle models. Although full dynamic models of the process are an attractive option to relate online signals to key performance parameters of the process, they are not considered to be typical for softsensors as the online solution of the models might be expensive. Moreover, the behavior of the solvers of dynamic equation systems for combination of noisy input signals is difficult to keep under control. Nevertheless, developments of online simulations of the processes are under way and are referred to as virtual plants. Versions that are able to simultaneously run with the process are used for process supervision (68).

69.4.19.1.9 Hybrid Models. A compromise between full dynamic process models and data-driven models can be made by combining the best of both approaches. From the full dynamic models, one can take the basic mass balance equations, which are simple to formulate, while the kinetic

69.4.19.1.8 Extended Kalman Filters. Simpler first principle models can be employed in softsensors when they are Process model dX F = m⋅ X − ⋅ X dt W dS F = −s ⋅ X + ⋅ (SF − S ) dt W dW =F dt dm =0 dt Xpred c pred

C est

Measurement model

OURpred = m ⋅ YOX ⋅ X + mOX ⋅ X CPRpred = m ⋅ YCX ⋅ X + mCX ⋅ X Basepred = YbX ⋅ ( XW − W 0X 0 )

Measurements

OUR, CPR, and base from on-line measurement signals of airflow,O2, CO2, and base consumption

OUR

= S pred

OURpred = H pred CPRpred

H = CPR

mpred

Basepred

Base

=

C pred

+

K

g

⋅

(H

pred

−

H)

Figure 69.9. Concrete scheme of an EKF estimating biomass and substrate concentrations as well as the specific biomass growth rate from OUR, CPR, and Base consumption measurement values.

CONCLUSION 10

40 Predicted Measured

OUR (g/kg/h)

Biomass (g/L)

20

10

6 4 2 0

0

5

10

15

0

5

10

15

10

7 Predicted Measured

6

Predicted Measured

8

5 CPR (g/kg/h)

Glucose (g/L)

Predicted Measured

8

30

0

1489

4 3 2

6 4 2

1 0

0 0

5 10 Fermentation time (h)

15

0

5 10 Fermentation time (h)

15

Figure 69.10. Typical example of EKF estimates of the biomass concentration X, the substrate concentration S, the OUR, and the CPR together with measured values.

submodels, which are much more difficult to establish, are represented in a data-driven way (e.g. by means of ANNs; Fig. 69.11.

69.5

CONCLUSION

A wide variety of sensor and instruments for process monitoring are available. Tremendous efforts have been

made in the recent years to enable PATs in bioprocesses. Several in situ and online technologies and sensor are on their way into commercial manufacturing to enhance processes and enable more advanced control systems. Most of them have been used in other industries for years and have evolved to meet the demands of the biotechnology industry. The application of new sensors, with the right selectivity, robustness, ease of use, and so on, will lead to enhanced process understanding and drive more advanced

Z −1

x cpr tai

m

ANN

x

dx =mx dt x

m tai

π

ANN

p/x

p dp =px dt p/x

Z −1

Figure 69.11. Scheme of an ANN-based approach of deriving the π (μ) relationship, that is, the product formation rate π as a function of the specific biomass growth rate μ.

1490


process controls as a PAT application. New and advanced sensors combined with strong data analysis tools will drive toward a next level of process control. Multivariate data analysis techniques and softsensors will provide a valuable toolkit in our efforts to better understand and monitor bioprocess and to improve their consistency. In general, the following trends can be seen in applications of softsensors: • an available sensor is not fast enough or not sensitive enough or displays adverse signal drifts; • a sensor allowing direct measurements at sufficient accuracy is not available; • a direct measurement is physically impossible; • the employment of a direct measuring device is too expensive; and • a direct measurement would require too much supervision or maintenance. The choice of the estimation algorithm depends on the accuracy and the computing time in online applications required for a particular task. For control purposes, smaller response times are mandatory, while for process supervision, usually more time is available. While the cost of sensing and computation has been decreasing drastically in the past, softsensor systems still require some development expenses. Hence, costs and benefits must carefully be balanced. Finally, it should be stressed that multivariate data mining techniques can discover patterns that are hidden within the process data sets and therefore lead to opportunities to increase robustness and yields. Software estimators do not generate new information. Hence, development and implementation of more accurate and reliable online sensors for production processes will be in important pursuit.

REFERENCES 1. Carloni A, Turner APF. Encyclopedia of industrial biotechnology. john wiley & sons; 2010. p 766–782 2. Nielsen J. Encyclopedia of industrial biotechnology. john wiley & sons; 2010. p 1–10. 3. Schepers TH, Lammers F. Curr Opin Biotechnol 1994; 5: 187–191. 4. FAD(http://www.fda.gov/cvm/guidance/published.html). 5. Arroyo AY. In: Baltz RH, Davies JE, Demain AL, editors. Manual of industrial microbiology and biotechnology. 3rd ed. ASM press 2010. 6. Olsson L, Schulze U, Nielsen J. Trends Anal Chem 1998; 17(2): 88–94. 7. Brooks SL, Higgins IJ, Newman JD, Turner APF. Enzyme Microb Technol 1991; 13: 946–955. 8. Ruzicka J, Hansen EH Flow injection analysis. 2nd ed. New York: Wiley; 1978.

9. Saucedo V, et al . Biotechnol Prog 2011; 27(3): 885–890. 10. Pons MN. In: Pons MN, editor. Bioprocess monitoring and control . Munich: Carl Hanser Verlag; 1991. p 4–28. 11. Stanbury PF, Whitaker A Principles of fermentation technology. Oxford: Pergamon Press; 1984. 12. Harris RJ, Shire SJ, Winter C. Drug Dev Res. Special Issue: Monoclonal Antibodies Drugs 2004; 61(3): 137–154. 13. Kjaergard L. In: Ghose TK, editor. Advanced biochemical engineering. New York: Springer-Verlag; 1977. p 131–145. 14. Pluschkel SB, Flickinger MC. Cytotechnology 1996; 19: 11–26. 15. Lee MS, et al . Biotechnol Prog 2003; 19: 1734–1741. 16. Sieblist C, Jenzsch M, Pohlscheidt M, Luebbert A. Underlying principles—bioreactor fluid dynamics. In: Moo-Young M, editor. Comprehensive biotechnology. 2nd ed, Volume 2. Elsevier 2011. p 47–62. 17. Pohlscheidt M, et al . Chem Ing Technol 2008; 80(6): 821–830. 18. Pohlscheidt M, et al . Vaccine 2008; 26: 1552–1565. 19. Hu WS, Aunins JG. Curr Opin Biotechnol 1997; 8(2): 148–153. 20. Aunins JG, Henzler HJ. In: Rehm HJ, Reed G, editors. Biotechnology. Stephanopolous G, Series editor. Volume 3, Bioprocessing. VCH publisher 1993. p 219–281. 21. Aehle M, et al . Biotechnol Lett 2011; 33: 2103–2110. 22. Kowollik S, et al . Chem Ing Technol 2010; 82(9): 1505–1506. 23. Pizarro SA, et al . Biotechnol Bioeng 2009; 104(2): 340–351. 24. Kao YH, et al . Biotechnol Bioeng 2010; 107(4): 622–632. 25. Lee YH, Tsao GT. Adv Biochem Eng 1997; 13: 35–58. 26. Atkinson B, Mavituna F Biochemical engineering and biotechnology handbook. 2nd ed. Chapter 1. Basingstoke, UK: MacMillan; 1991. 27. Carr-Brion KG. In: Carr-Brion KC, editor. Measurement and control in bioprocessing. New York: Elsevier Science; 1991. p 37–66. 28. Clark LC, Lyon C. Ann N Y Acad Sci 1962; 102: 29–45. 29. Reuss M. In: Finn RK, Prave P, editors. Biotechnology focus I: fundamentals, applications, information. New York: Oxford University Press; 1988. p 153–190. 30. Bean B, et al . Pharmaceut Manufact 2009; 9(1): 28–31. 31. Sonnleitner B, Locher G, Fiechter A. J Biotechnol 1992; 25: 5–22. 32. Konstantinov K, et al . Trends Biotechnol 1994; 12: 324–333. 33. Carvell J, Dowd J. Cytotechnology 2006; 50: 35–48. 34. Dowd J. Cytotechnology 2003; 42: 35–45. 35. Maskow T, et al . Biosens Bioelectronics 2008;; 24: 123–128. 36. Joeris K, et al . Cytotechnology 2002; 38: 129–134. 37. Broger M, et al . J Biotechnol 2011; 154: 240–247. 38. Lindberg W, et al . Cytometry 1993; 14: 230–236. 39. Derfus GE. Biotechnol Prog 2010; 26(1): 284–292. 40. Shaw A, et al . Appl Spectrosc 1999; 53: 1419–1428. 41. Lee H, et al . Vibrat Spectrosc 2004; 35: 131–137. 42. Noui L, et al . Chem Eng Process 2002; 41: 107–114. 43. Riley MR, et al . Biotechnol Prog 2001; 17: 376–378. 44. Scarff M, et al . Crit Rev Biotechnol 2006; 26: 17–39. 45. Schena M, et al . Science 1995; 270: 467–470.

REFERENCES

46. Charaniya S, et al . Biotechnol Bioeng 2010; 102(6): 1654–1669. 47. Kantardjieff A, et al . J Biotechnol 2010; 145(2): 143–159. 48. Gupta P, Lee KH. Trends Biotechnol 2007; 25(7): 324–330. 49. Xu X, et al . Nat Biotechnol 2011; 29(8): 735–742. 50. Hammond S, et al . Biotechnol Bioeng. 2012; 109(6): 1353–1356. 51. Flatman S, et al . J Chromatogr 2007; 848: 79–87. 52. Shukla A, Thömmes J. Trend Biotechnol 2010; 28(5): 253–261. 53. Larson T, et al . Biotech Prog 2003; 19: 485–492. 54. Chang Y, Bork C, Thömmes J. International Forum for Process Analytical Technology, Arlington, VA; 2005. 55. Chen L, et al . Control Eng Pract 2000; 8: 821–827. 56. Azevedo S, Dahm B, Oliveira FR. Comput Chem Eng 1997; 21: 751–756. 57. Schügerl K. J Biotechnol 2001;(85): 149–173. 58. Kirdar A, et al . Biotechnol Prog 2007; 23: 61–67. 59. Kourti T. Process Anal Technol 2005; 2(1): 24–29. 60. Dushesne C, et al . AIChE J 2002; 48(12): 2890–2901. 61. Kourti T. J Chemometr 2003; 17(1): 93–109. 62. Kourti T, et al . Comput Chem Eng 1996;(20): 745–750. 63. Kourti T. Process Anal Technol 2004; 1(1): 13–19. 64. Kourti T. Int J Adapt Control Signal Process 2005; 19: 213–246.

65. 66. 67. 68.

69. 70. 71. 72. 73. 74. 75. 76. 77. 78. 79. 80.

1491

Kourti. Annu Rev Control 2003; 27: 131–139. Sakoe H, Chiba S. IEEE Trans ASSP 1978; 26: 43–49. Nielsen N, et al . J Chromatogr 1998; A 805: 17–35. Kuprijanov K. Professional software toolbox advancing bioreactor performance. Running PhD Project, Martin-Luther-University, Halle-Wittenberg, Germany; 2011. Arndt M, Hitzmann B. Biotechnol Prog 2004; 20: 377–383. Chéruy A. J Biotechnol 1997; 52: 193–199. Eddy SR. Nat Biotechnol 2004; 22: 1177–1178. Jackson JE A user’s guide to principal components. New York: Wiley; 2003. Jenzsch M, et al . Bioproc Biosyst Eng 2006; 29(1): 19–27. Jolliffe IT. Principal component analysis. 2002; 2nd ed. New York: Springer. Kiviharju K, et al . J Ind Microbiol Biotechnol 2008; 35: 657–665. Lübbert A, Simutis R. Trends Biotechnol 1994; 12: 304–311. Luedeking R, Pieret EL. J Biochem Microb Technol 1959; 1: 393–412. Simutis R, Lübbert A. Biotechnol Prog 1997; 13(4): 479–487. Johnson KC, et al . Biotechnol Bioeng 2011; 108(2): 475–480. Charaniya SC, et al . Trends Biotechnol 2008; 26(12): 690–699.