SE518174C2 - Detection method of unwanted microbial infection in animal cell culture - Google Patents

Abstract

A method for early detection of undesired microbial infection in a cell culture by means of an electronic nose having an array of gas sensors and the ability of pattern recognition is described, said method comprising the steps of withdrawing an off-gas sample containing volatile compounds from the cell culture in a bioreactor, passing said sample to the gas sensors of the electronic nose, and comparing any specific gas sensor signal patterns with standard signal patterns corresponding to different micro-organisms, wherein the off-gas sample is withdrawn from an animal cell culture, and the detection of any microbial infection is established about 1-50 h earlier for each micro-organism compared to known techniques.

Description

25 30 35 ø n | ; now n 513 174 2 complicated sampling and evaluation, despite the costly arrangements.

A few recent observations suggest that electronic noses could be attractive non-invasive alternatives for the detection of microbial contamination in bioreactors for the cultivation of microorganisms (1998). The principle of electronic noses 1994) is well known from several to- (Namdev et al, (Gardner and Bartlett, applications for detection of components in various types of samples and media by detection of emitted gases).

The detection is based on the registration of the emitted gas from cultures by means of sets of gas sensors which are sensitive to various volatile compounds, eg hydrocarbons, such as aromatic compounds (eg ergosterol), aldehydes (eg acetaldehyde) and alcohols. (eg ethanol); sulfur-containing compounds, nitrogen-containing compounds, hydrogen and carbon oxides, which are specific to each bacterial strain. By simply extracting a continuous sample stream of a gas discharged from a bioreactor, a non-invasive almost real-time monitoring of the release is achieved.

Since different strains of organisms give rise to specific gas sensor signal patterns, electronic noses can now constitute a new type of instrument for early detection of different types of microbial contamination of cultures (Craven et al, 1994; Holmberg et al, 1999). The fact that the electronic noses enable measurement outside the sterile barrier of the bioreactor system is in line with the requirements for industrial production and the ease of adaptation for real-time monitoring and as operator support.

WO97 / 08337 (Unipath Ltd) describes an electronic nose for the detection of microorganisms in blood cultures inoculated with microorganisms.

Gebrauchsmuster DF299 02 593 Ul describes a gas analyzer for medical diagnostics. 10 15 20 25 30 35 513 174 if? 3 There has long been a clear need for a faster method for the detection of microbial infections in bioprocesses, especially in animal cell cultures.

However, it has been recognized by those skilled in the art that an electronic nose would not be advantageous over conventional techniques for detecting microbial infections in animal cell cultures, as it seems unlikely that detection could be possible at such an early stage of the microbial contamination, ie that the electronic nose would probably not give any response at all or a very weak or diffuse response.

As regards the detection of the product concentration in animal cell cultures for the production of biopharmaceuticals, the use of such cultures has steadily increased over the past three decades. After initially being based on ascitic production in vivo or production with a roller bottle since the mid-1980s, the emphasis has now been on the development of highly efficient animal cell processes [16]. According to this trend and taking into account the fact that direct measurement in animal cell culture is still limited to a small number of environmental conditions [l2], the development of sophisticated monitoring methods has become a necessity for precise process control.

One of the complications of efficient production of recombinant protein in a biological system is the lack of rapid direct methods for the detection of the desired product. Today's analytical methods are often too slow for efficient closed-loop control, which is why protein production must be controlled via predetermined operating profiles.

Conventional product monitoring in animal cell cultures often relies on immunoassay methods that are usually used only offline for documentation purposes instead of directly. Examples of later advances in bioproduct analysis are those (on-line) for process control [6] 10 15 20 25 30 35 o ø -. successful use of capillary electrophoresis [6], an optical biosensor PH and direct immunoassay [15]. However, despite their sensitivity and accuracy, these methods must still overcome the problem of sterile and reliable in situ sampling from the bioreactor.

The problem of aseptic sampling becomes particularly prominent when monitoring the outflow from the bioreactor's gas space.

OBJECTS OF THE INVENTION An object of the present invention is to eliminate the above-mentioned problems involving slow detection of microbial infections in animal cell cultures. Another object of the present invention is to provide a method for estimating the product concentration in a cell culture.

These objects have been achieved by means of methods of the kind mentioned in the introduction, which have the features stated in claims 1 and 6. Additional features are defined in the subclaims.

Description of the drawings Fig. 1 (a) shows time profiles for selected sensor response signals from the electronic nose and culture parameters for an rFVIII culture in production scale. A bacterial infection occurred at the end of the culture. The MOSFET 3 response is shown magnified in its measurement interval of 20 minutes during the infection period. The first reaction of the pO2 probe to the infection is indicated in the curve. The cultivation parameters were normalized between 0 and 1. The sensor signals are given in arbitrary units.

Fig. 1 (b) shows time profiles for selected sensor response signals and culture parameters for a successful production scale rFVIII culture. The culture parameters were normalized. The sensor signals are specified in arbitrary units.

Fig. 2 (a) shows time profiles for selected sensor response signals and culture parameters for a laboratory scale rFVIII culture. The culture was contaminated with 10 15 20 25 30 35. »A ~ u 'n 518 174 5 Bacillus cereus. F1, F2: introduced process errors via a reduction of dissolved oxygen to 0% saturation. M: the electronic nose was disconnected from the process. The MOSFET 3 response is displayed magnified in its measurement interval of 20 min during the infection period. The first reaction of the pO2 probe to the infection is given in the curve. The cultivation parameters were normalized. The sensor signals are given in arbitrary units.

Fig. 2 (b) shows time profiles for selected sensor response signals and culture parameters for a laboratory scale rFVIII culture. The culture was contaminated with Pseudomonas aeruginosa. The MOSFET 3 response is shown magnified in its measurement interval of 20 minutes during the infection period. The first reaction of the pO2 probe to the infection is given in the curve. The culture parameters were normalized. The sensor signals are specified in arbitrary units.

Fig. 3 (a) shows time profiles for selected sensor signals and culture parameters for a control fermentation on a laboratory scale. Pseudomonas aeruginosa was added to a reference culture medium at the indicated time. The culture parameters were normalized. The sensor signals are expressed in arbitrary units.

Fig. 3 (b) shows time profiles for selected sensor signals and culture parameters for a control fermentation on a laboratory scale. Bacillus subtilis was added to a reference culture medium at the indicated time.

The culture parameters were normalized. The sensor signals are specified in arbitrary units.

Fig. 4 (a) shows the original and adjusted gas sensor signal from a production scale rFVIII culture.

The sensor signals are specified in arbitrary units.

Fib 4 (b) shows the time profiles of selected culture parameters and gas sensor signals for a laboratory-scale rFVIII culture. Commercial interests require a normalization of the cultivation parameters between 0 and 1. The sensor signals are expressed in arbitrary units. Fig. 5 shows the characterization of the gas sensor behavior in case of process parameter changes. The sensor signal is specified in arbitrary units.

Fig. 6 shows time profiles for selected sensor signals from the electronic nose and important culture parameters for an rFVIII culture in production scale. The cultivation parameters are normalized between 0 and 1. The sensor signals are specified in arbitrary units.

Fig. 7 shows the estimated and measured concentration of recombinant factor VIII in relation to the culture time in a rFVIII culture on a production scale. The estimated values were obtained from a neural network model trained with data from an electronic nose from 5 rFVIII cultures on a production scale. The rFVIII concentration was normalized between 0 and 1.

Fig. 8 shows the estimated and measured number of viable cells in relation to the culture time from a rFVIII culture on a laboratory scale. The estimated values were obtained from a neural network model trained with data from an electronic nose from a laboratory-scale rFVIII culture. The number of viable cells was normalized between 0 and 1.

Summary of the Invention The present inventors have surprisingly found a method for early detection of microbial infection in an animal cell culture process using an electronic nose. Infections with defined cell concentrations have also been established to characterize the electronic nose's response to three infectious organisms in particular that are common in bioprocesses.

The present inventors have furthermore found a method for detecting the product concentration in animal cell cultures by means of an electronic nose which includes an artificial neural network and which has the ability to pattern recognition, which has not been previously known or possible. In addition, both of the above-mentioned methods developed by the inventors can be carried out simultaneously in the same cell culture process by means of the same electronic nose.

The terms "microbial" and "microorganism" as used herein refer to any unwanted bacteria, yeast species and fungi that could infect or contaminate the bioreactor contents.

The cell culture in which potential microbial contamination or infection is to be detected is preferably an animal cell culture, for example a human or animal cell culture, for example a fibroblast cell culture, a hamster cell culture, a hybridoma cell culture, an insect cell culture, etc.

The cell culture medium can be any conventional medium used in this field. The volume of the cell culture medium can also vary from laboratory scale up to large scale (0.5 - 20,000 l).

The gas emitted from the cell culture accumulates in the gas space above the cell culture surface of the bioreactor and includes volatile compounds, such as hydrocarbons, which are emitted by and are specific for any microorganism in the culture.

In a preferred embodiment of the present invention, a culture of recombinant CHO cells (Chinese hamster ovary cells) producing human factor VIII (rFVIII) is detected with respect to microorganisms, and tests have been performed in a suitable medium for 2 hours. - liter- and 500-liter scale. After an initial batch addition phase, the culture is continuously flushed with fresh medium using an external cell retention device. Samples were taken every 24 hours, and cell concentration and viability were measured by exclusion using dye.

The term "electronic nose" as used herein refers to a device containing a set of gas sensors that provide a characteristic response to a gaseous sample, the device also enabling pattern recognition. 10 15 20 25 30 35 ø o u | co 518174 8 The term "artificial neural network" as used herein refers to a system used in a computational method based on the theory of artificial neural network, which is described by, for example, C M Bishop, Neural Networks for Pattern Recognition, Oxford University Press 1995.

The term "pattern recognition" as used herein refers to the identification of a pattern of responses with known variables using any calculation and / or non-calculation method. The non-calculation method is, for example, visual observation of patterns of sensor signals.

The preferred electronic nose used in connection with the present invention is well known (NST3310) and is available from Nordic Sensor Technologies AB, Teknikringen 8, Mjärdevik Science Park, 583 30 Linköping, Sweden. It also includes an artificial neural network system in the embodiment that involves estimating the product concentration. The electronic nose sensor set used in the 500 l rFVIII culture was in its preferred embodiment equipped with 10 MOSFET sensors (metal oxide semiconductor field effect transistors, with catalytic metal controls of palladium, iridium or platinum, operated at different temperatures (140/170). ° C), 12 MOS sensors (semiconductor metal oxide sensors of the Taguchi or FIS type; SnO2 sensors operated at 400 ° C) and 1 CO2 sensor based on infrared adsorption.

In the aforementioned 2 L rFVIII cultures, the sensor set contained 10 MOSFET sensors, 19 MOS sensors and 1 CO2 sensor based on infrared adsorption. A mass flow regulator maintained a stable flow across the sensors.

The sensors were arranged in a gas flow injection system where the gas flow first passed the MOSFET sensors and then the MOS sensors and the C02 sensor. The electronic nose was connected outside the sterile barrier to the bioreactor line for the specified gas. Sampling was performed continuously from the gas discharged from the reactor for usually 30 seconds every 20 minutes throughout the process.

The types and number of sensors in the electronic nose depend on the composition of the culture medium and the organisms in the medium to be detected. Thus, a varying number of MOSFET, MOS and CO 2 sensors and various combinations thereof can be used for the detection of microorganisms.

Other sensors that could be useful are quartz crystal, conductive polymers and / or semiconductor sensors.

The gas sensor responses from the rFVIII cultures showed different signal shifts. The shifts were removed using an algorithm that subtracts / adds the remaining offset from a sensor signal shift.

Control fermentations were performed in a 3 l bioreactor system (Chemoferm AB, Sweden) which was equipped with a process control unit of type KENT-B2® (Belach Bioteknik AB, Sweden). The culture parameters were set to conform to the rFVIII culture conditions. The gas flow was kept constant at 800 ml / min of compressed and filtered air throughout the process.

A quadropole mass spectrometer was connected to the bioreactor line for emitted gas (Multi-Gas, Leda-Mass Ltd, Cheshire, UK). The instrument was equipped with a dual detector system (faraday and electron multiplier detector).

The gas emitted from the bioreactor was screened for atomic weights 1-300 at maximum refresh rate (about 5 minutes).

The electronic nose connected to the bioreactor system was the same instrument used for the 2 l rFVIII cultures described above.

Any microorganism, such as a bacterial strain, which contaminates or infects a cell culture medium can be detected by the electronic nose in the method of the present invention, but in addition to 518 174 Bacillus cereus and Pseudomonas aeruginosa are detected.

The Bacillus cereus and Pseudomonas aeruginosa bacterial strains used for infection studies in the 2 L rFVIII cultures were isolated by Pharmacia & Upjohn AB.

The strains used in the control fermentations were Pseudomonas aeruginosa (isolated at Linköping University Hospital, Linköping, Sweden), Bacillus subtilis (CCUG l63B) and Micrococcus luteus (CCUG 5858). LB medium (1% tryptone, 0.5% yeast extract and 1% NaCl) was used to maintain the strain. Control fermentations were performed in a defined serum-free Eagle's MEM-like cell culture medium (Shuler et al, 1992).

The number of colony forming units (cfu) was determined by colony counting after incubation of media samples on agar plates. Copies of media samples were applied to LB aga plates and incubated at 30 ° C for 24-48 hours.

Thus, with the present invention, regardless of the microorganism to be detected, the detection can be performed for several hours, i.e. about 50 hours, usually about 1-10 hours, and most preferably 5 hours, earlier for each microorganism than in comparative methods, which is by special interest in large-scale bioprocesses. Different microorganisms can be detected at different times in their growth phase, but in general the detection is possible within the above-mentioned time intervals earlier than is possible with conventional methods. As for the product concentration estimation embodiment, several types of products can be detected by the method of the present invention, preferably proteins, but also other biopolymers, such as polycarbonates and polynucleotides.

It is surprising that gaseous compounds in the exhaust gases emitted from bioreactors and derived from certain products in the culture medium can be detected. This embodiment regarding estimation is not limited to animal cell cultures, but also to microorganism cell cultures and other types of media. Preferably, this embodiment of the present invention allows direct estimation of the factor VIII concentration in production scale cultures.

Experiments and results for the embodiment regarding detection of microbial infection Fig. 1 (a) shows how it is possible to trace the occurrence of bacterial infection in a recombinant factor VIII culture of 500 l based on the response signals of the electronic nose. The three sensor signals in the diagram show the monitoring of the culture for a period of 18 hours in parallel with the measured number of viable cells, the rFVIII concentration and the total flow rate through the bioreactor. The signals reflect the fundamental changes in the product titer and the concentration of viable cells, as reported in detail in other 1999b). On the seventeenth day of culture, a technical defect occurred in the plant, which was thought to have caused a bacterial contamination site (Bachinger et al., In the culture. After 16 hours, the contamination was confirmed by a reduction in the concentration of dissolved oxygen in the medium, which was monitored with the DO electrode. Later, the infection was verified, and the infecting organism was identified as Bacillus cereus.

As shown in Fig. 1 (a), the response from the electronic nose indicated the infection as early as 6-8 hours after the process defect had occurred. The figure also shows that the number of viable cells and the flow rate were stable during this period. A comparison with a rFVIII culture of 500 l shows the clear differences between the sensor signal patterns from both cultures (Fig. 1 (b)).

The electronic nose is thus able to indicate the bacterial infection about 10 hours before the DO sensor. It should be noted, however, that the sensors do not explicitly reveal the reason for the signal deviation and infection. 10 15 20 25 30 35 _1 cfu / ml medium. 5, 8 1,4 = fr: ua 12 In order to further evaluate the potential of electronic noses for rapid detection of contamination in animal cell culture processes, three frequently occurring bacterial strains were intentionally added to test cultures.

First, Bacillus cereus was added to a 2 L scaled-down rFVIII culture at a concentration of Fig. 2 (a) showing the response signals of the electronic nose. 8 hours before the addition of the bacteria, a process defect was introduced at the time marked F2 in the diagram. The concentration of dissolved oxygen was manually reduced to 0% over 30 minutes, and the sensor signals immediately indicated the process deviation (see Fig. 2 (a)). However, the sensor signals did not recover, but instead decreased gradually after this incident.

When the same process defect was introduced during a period 10 hours earlier in the culture (F1), the sensors immediately reached the original signal level after pO2 was reset. One explanation for this behavior may be that the cells after almost three weeks of culture were more vulnerable than at the beginning of the culture. The gas sensors monitored the cellular activity, which gradually decreased, indicating that the cells did not recover properly after being exposed to low oxygen concentrations. The sensor signals were suddenly stabilized about 1 hour after the bacteria were added to the cell culture. This took place 9 hours before the pO2 sensor detected a decrease in the concentration of dissolved oxygen.

Second, Pseudomonas aeruginosa was added at a concentration of 0.1 cfu / ml medium (Fig. 2 (b)) to the same culture and reactor system. No signal response change was noted after addition of the bacteria. A signal deviation from the original pathway could not be detected until 15 hours after the infection. The pO2 sensor identified the process deviation 6 hours later.

In order to further verify the validity of the electronic nose signals for the Bacillus and Pseudomonas species, these were incubated in the same medium, but without the CHO culture. Figure 3 (a) shows the response of the electronic nose to a control fermentation in which 100 colony forming units (cfu) of Pseudomonas aeruginosa per ml of cell culture medium were inoculated. 1 h after inoculation, the contamination was indicated by a response change in the selected MOSFET sensor. 14 hours after inoculation, the pO2 sensor indicated a decrease in dissolved oxygen content. In parallel, an in-line mass spectrometer was used in this culture to monitor the gas emitted from the bioreactor. The recorded partial pressure for CO 2 shows that the mass spectrometer does not indicate the infection earlier than the pO2 sensor.

The response of the electronic nose to a control fermentation in which 4000 cfu / ml of Bacillus subtilis was inoculated is shown in Fig. 3 (b). This large number of cells caused an immediate signal change in one of the sensors plotted in Fig. 3 (b). The pO2 sensor monitored a decreasing level of dissolved oxygen 7 hours after inoculation. The partial pressure of CO2 recorded by the mass spectrometer also did not indicate the infection earlier than the pO2 sensor.

By repeating the cultures in cell-free media with P. aeruginosa and B. subtilis at lower concentrations (0.5 cfu / ml and 20 cfu / ml, respectively), expected longer response times for the electronic nose were demonstrated.

An infection of cell-free medium with Micrococcus luteus verified our previous observations in 500 l where the electronic nose responded with a clear signal deviation for a contamination that was later identified as Micrococcus sp. There we had reported the identification of the bacterial infection using the electronic nose about 2 days before the pOš signal indicated contamination. During this control fermentation, it was found that the pO2 sensor indicated growth 39 hours after inoculation. A signal change from the electronic nose could be observed 10 hours after inoculation.

In Table I, the results of the contamination tests are summarized. 10 15 20 25 30 35 518174 o n uoovoc 14 Thus, in particular, the response from an electronic nose equipped with semiconductor metal oxide sensors on flushed CHO cultures producing recombinant human factor VIII was examined. scale, bacterial infections of recombinant factor In production scale and laboratory VIII cultures could be identified earlier than by a conventional method, i.e. an electrode for dissolved oxygen.

Accidental and controlled addition of bacterial contaminants to factor VIII cultures resulted in a clear sensor response pattern change approximately 5-10 hours earlier than the expected reduction in dissolved oxygen content. However, the electronic nose will not explain the reason for the indicated pattern change. It is therefore more realistic in this specific application that the instrument only serves as a support for the operator.

The regulated addition of a small number of cells of Pseudomonas and Bacillus strains showed the high sensitivity of the electronic nose. This result also confirms that the instrument is a sensitive biomass monitoring unit.

Experiments and results for the embodiment regarding UDD estimation of product concentration Experiments regarding the estimate of the product concentration in a cell culture were performed with the same equipment for the electronic nose as in the previous embodiment regarding detection of microbial infection, but were supplemented with a neural network and were also in the scale 2 l and 500 l respectively.

The average cultivation time was about 5 weeks.

Samples were taken every 24 hours, and cell concentration and viability were measured by exclusion using dye.

The lactate concentration was analyzed using a Model 2700 YSI analyzer (Yellow Springs Instruments, Yellow Springs, OH). The concentration and biological activity of rFVIII were determined using approved methods. 10 15 20 25 30 35 ¿._, .._.__. _, _ _: ____: 518 174 = "=" - = - =: = -s s. '= -: en =' - u. - u n. u n ..... .. 0 Z 15 The distance between the electronic nose and the bioreactor outlet was 3 m.

The bioreactor system was arranged in such a way that it complied as much as possible with the conditions for production scale. Samples were taken in varying time intervals every 12 or 24 hl Analyzes outside the system were performed as described in section 2.1.

The electronic nose was equipped with 10 MOSFET sensors, 19 MOS sensors and 1 C02 sensor. In addition, a mass flow control device was introduced into the flow injection system to establish a stable gas flow across the sensors. The sample interface described under 2.1 was modified in such a way that the distance to the bioreactor outlet was further reduced to 0.7 m.

The electronic nose used in the production scale crops was also incorporated into the monitoring equipment.

The sensor shift exclusion algorithm and the median filter were developed in Object Pascal (Delphi 4, Inprise Corp., CA, USA). The shift exclusion algorithm subtracts / adds the remaining offset from a gas sensor signal when the shift exceeds a certain percentage and signal level and progresses in only one direction.

The implementation of the algorithm for direct conversion should be uncomplicated. The median filter was applied to data from a shift exclusion sensor with a window size covering five consecutive measurement points.

The programming language MATLABTM (The MathWork Inc., MA, USA), supported by MATLABTM toolboxes for neural networks and chemometrics, was used for performing calculations with the artificial neural network (ANN) and in main component analysis (PCA). The PLAT toolbox for MATLABTM was used for the component correction method and the forward selection procedure (Eigenvector Technologies, Manson, WA).

Principal Component Analysis (PCA) is a linear unattended pattern recognition technique that can be used for reduction 10 15 20 25 30 35 ø n | o v: o n 518 174 ä? 16 of the dimensionality of multivariate data BJJ. The calculation results in main components that describe the direction of the variation in the data set. The results are presented in diagrams where the two most important main components are plotted against each other, either as a score plot, which describes how the samples are related to each other, or as a "loading plot", which describes the relationships between the variables.

The neural network structure used in this work was a standard hidden layer backpropagation ANN (17N) with a sigmoidally activated function and the Levenberg-Marquardt update algorithm [14] The network structure for estimating the rFVIII concentration 11 nodes in the hidden layer, 4 input signal variables and 1 output signal. The minimum error gradient was 0.000l, the initial value for Ä was 0.001 and the multipliers for increasing and decreasing Ä were 10 and 0.1, respectively. The error rate was set to 30, and 20 practice cycles were required to achieve optimal performance for the model.

To estimate the number of viable cells, the ANN equipment consisted of 9 hidden nodes, 6 input signals and 1 output signal. The minimum error gradient was 0.0001, the initial value for Ä was 0.001 and the multipliers for increasing and decreasing Ä were 10 and 0.1, respectively. The error rate was 0.5, and 30 practice cycles were performed.

In the component correction method (CC, DJ), a filter is used to reduce sensor operation in order to increase the service life of pattern recognition models. CC assumes that the operation is not randomly distributed, but instead has a preferred direction in the measuring space. A PCA loading vector is calculated on the basis of the calibration measurements, which capture the direction in the response space. Subtraction of the data set from the original culture from the projection of the culture values on this charge vector results in the 1018 20 25 30 35 518 174: -; f .. = t »1" * 17 desired the elimination of the direction describing the operation. This preserves all other directions and significant variances that separate accumulations and concentrations. A trend line was adjusted by linear regression to the mean of the calibration measurements for each sensor signal covering all performed cultures.

The culture values were mean-centered when performing the component correction method with adapted calibration data.

Details of the forward-looking selection procedure are given by Eklöv et al [5]. The purpose is to find a subset of original sensor signals that minimizes a selection criterion. The selection criterion is the predictable error from a model of multiple linear regression in a direction towards the desired model yield (process variable). A forward-looking selection adds one variable at a time to the model until the selection criterion reaches a minimum. The selected sensor signals contain relevant information for variable estimation, which thus represents satisfactory parameters for use as input signals to the artificial neural network.

The culture data sets were processed with the shift exclusion algorithm, filtered and calibrated before a search was performed for correlation against rFVIII concentration and the number of viable cells.

The electronic nose's multi-signal response enables direct estimation of the main parameters through the use of pattern recognition models. Theoretically, a model should be made up of a maximum number of data sets to ensure that it covers the entire variation of the process. This and the degree of correlation of the signals with respect to the estimated parameter 10 15 20 25 30 35 51 3 1 7 4 Ü ÃÃ-åi-.Ií- 18 determine the models, the flexibility and the reliability.

In the present study, 6 factor VIII (rFVIII) cultures were monitored on a production scale with the electronic nose. Each of the cultures took place for 5 weeks, and the total monitoring time was extended by more than 5,000 hours of the direct operation. Therefore, a significant part of the variation of the process was covered, as well as the instrument performance and calibration.

A representative gas sensor signal from one of the cultures is shown in Fig. 4 (a) (see MOSFET response). Sudden response shifts can be observed in the signal. The use of multivariate methods for parameter estimation is made impossible by this phenomenon.

In order to better understand these shifts and validate the response characteristics of the gas sensors, the rFVIII cultures were scaled down to 2 l and monitored in parallel with two electronic noses.

A major question in interpreting these signal changes was whether they were the result of CHO cell metabolism, the production scale bioreactor set, or the instrument itself. As can be seen in Fig. 4 (b), similar signal shifts recurred in the scaled-down cultures (see MOSFET 2 response). This excludes the design and dimension of the bioreactor on a production scale as a cause.

Instrument defects can also be ruled out because both electronic noses showed the same response characteristics.

We therefore hypothesized that the biological activity of the CHO cells during the production of rFVIII caused the signal shifts. In addition, during long-term culture of a VERO cell line with flushing, the shifts were not noted.

In order to more fully understand the behavior of the gas sensor, it is important to identify the non-biological parameters in the process that have a direct impact on the gas sensors' response. Measurements on a reference culture medium showed that changes in temperature and concentration of dissolved oxygen resulted in a permanent re- 10 15 20 25 30 35 518 174 ä? Iå-åß Iš. 19 sponsorship change for metal oxide sensors (Fig. 5). It has previously been shown that a change in the aeration speed has no direct effect on the sensor response D3].

Operating parameter changes of this type often occur in rFVIII cultures, and the resulting signal shifts should therefore be eliminated from the sensor signals.

The observations presented above also indicate that it is a correct approach to remove the shifts that we presume are caused by the cells. Thus, an algorithm was developed that enabled controlled elimination of the signal offsets. An adjusted sensor response signal from a production scale rFVIII culture is shown in Fig. 4 (see MOSFET shift eliminated response).

The operation of solid state gas sensors is a major problem for electronic nose technology [8]. Established pattern recognition models would require constant recalibration in the event that no operational mitigation is used.

Especially for long-term bioprocesses, this would be time consuming and inefficient.

One possibility would be to repeatedly recalibrate the sensors with a suitable reference gas or gas mixture. The reference gas mixture should not only be correlated to the measured sample, but also the operation of both should be highly correlated. In a bioprocess, however, one must theoretically first identify the most important volatile compounds produced during the fermentation process and use these as a reference. However, this would be impractical, and it is therefore better to find a common substance that is present in high concentrations throughout the process.

The calibration of the gas sensors was realized by recording their response to the volatile compounds from the culture medium in the bioreactor shortly before inoculation. At this stage, the gas space mainly contains water vapor. Water vapor has previously been found to be suitable for 10 15 20 25 30 35 ø - | - .v 513 174 Ü -.I = '' z 20 calibration purposes in several applications on food samples [9]. We have experienced that this calibration procedure provides a correct representation of the operation of gas sensors that are exposed to a bioprocess.

Twenty calibration measurements were performed shortly before the inoculation of each of the cultures. The response for each sensor was recalculated to a mean, and a trend line was adjusted by linear regression covering all performed cultures. Operation over a period of 1 year for selected sensors is listed in Table 2. Several anti-operational methods have been developed in recent years [10,9,1]. In this study, the component correction (CC) method Ü] was applied to the sensor signals, as described in section 2.4. Sensor operation, which for some sensors amounted to 30% for 1 year, was removed with great success using this method.

In Fig. 4 (b), several sensor signals from the scaled-down rFVIII culture are plotted. The higher signal variation compared to that in the production scale was a result of installation of the instrument near the reactor outlet. The signals were processed with a connected averaging filter to reduce noise and facilitate visualization.

State variables and selected sensor signals from a production scale production are shown in Fig. 6. The plotted signals were processed using the above-described shift exclusion algorithm and represent a selection of the signal variation among the 110 different signals derived from the gas sensor arrangement.

Some of the sensor response signals in Fig. 6 show correlation with the rFVIII concentration in the bioreactor.

The deviation signal, on the other hand, follows the increase in specific lactate formation at the end of the cultivation. The correlation with the number of viable cells is evident from some of the signal responses. However, the resolution of the sensor signals is not high enough at the beginning of the culture for an accurate direct estimate of the cell number. This was due to 10 15 20 25 30 35 ø ø u. av u n 513 174 É- * Éïïš Ü Én * = u u nu. u. 21 probably on the long distance between the reactor outlet and the electronic nose which was necessary due to the test set-up and the instrument dimension in production scale. In the scaled-down rFVIII cultures, it was possible to dissolve small biomass changes even at inoculum cell densities due to the reduced sampling set (see Fig. 4 (b)).

The gas sensor signals from all six cultures on a production scale, which were treated with the signal elimination algorithm, where the noise of the signals was partially removed by using a median filter, and with the component correction calibration were used to estimate the rFVIII concentration.

The forward selection method described in the method section was used to select the signal parameters with the highest correlation with the rFVIII concentration. The selected signals from five cultures on a production scale were compiled, and a neural network was trained with the product concentration measured outside the system. The selected sensor signals were MOSFET 3-off derivatives, MOSFET 1-off derivatives, MOSFET 8-response and MOSFET 3-response. The topology of the neural network was identified using "trial and error".

The model with a preserved neural network was validated against a culture that was unknown to the model. The result of the validation is shown in Fig. 7. The total estimation error was about 10% with an average deviation of 1.1 IU / ml. With an rFVIII peak value that varied by as much as 50% between the six cultures, this is still an acceptable result. It also suggests that the use of additional culture data sets for the ANN model training would increase the accuracy of the estimate.

To enable sophisticated control of recombinant protein production, it is important to monitor the number of viable cells. Although the sensor signal resolution in production scale cultures was not high enough for accurate direct estimation of the number of viable cells, 51 g 174 Ü 'E o .. u 22 scaled-down rFVIII cultures allowed this. The values from only two cultures are not sufficient to cover the variation of this process. The resulting model will therefore not have the same impact as the model established for estimating product concentration. V A The sensor responses were processed with the above signal exclusion algorithm and filtered with a median filter. one was found to be marginal (about 2%). Sensor signals with No sensor calibration were performed, as operating correlation with the number of viable cells was selected using the forward-looking selection procedure.

MOSFET 1 response, MOSFET 3 response, Taguchi 2 response, FIS 2 response, FIS 5-off derivatives and MOSFET 2 "off-integral" were identified by the selection method.

The selected sensor signals from both cultures were pooled, and 60% of the values were used to train an artificial neural network against the number of viable cells. An appropriate neural network topology was identified using trial and error. The final ANN model was validated against the remaining 40% of the data set. An average deviation of 0.21 x 105 cells / ml shows that the correlation with the number of viable cells is found in the sensor signals.

To obtain a representative model for estimating the number of viable cells, a neural network was trained with sensor responses from one of the cultures. The same selected signals were used in the model described above. Fig. 8 shows the validation result of the trained model against the second rFVIII culture. The total error in estimation was about 10% with a mean deviation of 0.41 x 10 6 cells / ml.

The use of an electronic nose for direct estimation of the product concentration in a flushed culture of recombinant animal cells was thus shown here.

It was shown that the factor VIII concentration in CHO cell cultures on a production scale could be estimated at 10 | . - Q n. 518174 nu vv 23 satisfactory accuracy by combining the electronic nose with artificial neural network technology.

The results confirm that the multivariate response of the electronic nose contains considerable information about the product concentration in this particular process. It can be assumed that this is related to cell stress caused by excessive expression of the recombinant protein. It should therefore be possible to estimate the product concentration in any recombinant process.

The possibility to directly estimate the number of viable cells was also presented. In laboratory-scale factor VIII cultures, we achieved an acceptable accuracy for estimating cell concentration. 10 l5 20 25 30 35 40 45 - nan nu 518174 24 REFERENCES Bachinger, Th., Riese, U., Eriksson, R., Mandenius, C.F. (1999a) Monitoring cellular state transitions in a production-scale CHO-cell process using a chemical multisensor array. To appear in J. Biotechnol.

Bachinger, Th., Riese, U., Eriksson, R.K., Mandenius, C.F. (1999b) Estimation of product concentration in a perfused animal cell cultivation using an electronic nose. (as a manuscript).

Craven, M.A., Hines, E.L., Gardner, J.W., Horgan, P., Morgan, D., Ene, I.A. (1994) Bacteria detection and classification using artificial neural networks in conjunction with an electronic nose. Neural networks and expert systems in medicine and health care, Plymouth, UK, 1994, pp. 226-234.

Elmroth, I., Valeur, A., Odham, G., Larsson, L (1990) Detection of microbial contamination in fermentation processes: Mass spectrometric determination of Leuconostoc mesenteroides cultures. Biotechnol. Bioeng. 35, 787-792.

Gardner, J.W. and Bartlett, P.N. (1994) A brief history of electronic noses. Sensors and Actuators B 18-19, 211- 220. (1999) Exploiting antibody-based technologies 290-296.

Harris, B. to manage environmental pollution. TIBTECH 17, Holmberg, M., Gustafsson F., Hornsten E. G., Winquist F., Nilsson L.E., Ljung L., Lundstrom I. (1998) Bacteria classification based on feature extraction from sensor data. Biotechnol. Techn. l2 (4), 319-324.

Namdev, P.K., Alroy, Y., Singh, V. (1998) Sniffing out trouble: Use of an electronic nose in bioprocesses.

Biotechnol. Prog. 14, 75-78.

Shuler, M.L., Kargi, F. (1992) Bioprocess considerations in using animal cell culture. In: Shuler, M.L., Kargi. F. (ed.), Bioprocess engineering: basic concepts. Prentice Hall PTR, New Jersey. 10 15 20 25 30 35 40 45 25 1) Artursson, T., Eklöv, T., Lndström, I., Sjöström, M., Mårtensson, P., Holmberg, M .. Drift correction for gas sensors using multivariate methods. Submitted (1999). 2) Bachinger, Th., Mårtensson, P., Mandenius, C.F .: On-line estimation of biomass and specific growth rate in a recombinant Escherichia coli batch cultivation process using a chemical multisensor array. J. Biotechnol. 60 (l998a) 55-66. 3) Bachinger, Th., Lidén, H., Mårtensson, P., Mandenius, C.F .: On-line estimation of state variables in baker's yeast fermentation using an electronic nose. Sem.

Food Anal. 3 (l998b) 85-91. 4) Bachinger, Th., Riese, U., Eriksson, R.K., Mandenius, C.F .: Monitoring cellular state transitions in a production-scale CHO-cell process using an electronic nose. To appear in J. Biotechnol. (1999). 5) Eklöv, T., Mårtensson, P., Lundström, I .: Selection of variables for interpreting multivariate gas sensor data. Anal. Chim. Acta 381 (1999) 221-232. 6) Freitag, R., Reif, O.-W., Weidmann, R., Kretzmer, G .: Production of recombinant h-ATIII with animal cell cultures using capillary electrophoresis for product monitoring. Cytotechnol. 21 (1996) 205-215. 7) Gill, A., Bracewell, D.G., Maule, C.H., Lowe, P.A., Hoare, M .: Bioprocess monitoring: An optical biosensor for rapid bioproduct analysis. J. Biotechnol. 65 (1998) 69-80. 8) Göpel, W., Schierbaum, K.D .: Definitions and typical examples. In Göpel, W., Jones, TA, Kleitz, M., Lundström, I. and Seiyama, T. (eds.), Chemical and biochemical sensors, del I, 2, sid 1-28, VCH Verlagsgesellschaft, Weinheim, Germany (1992). 9) Haugen, J.E., Tomic, O., Kvaal, K .: A calibration method for handling the temporal drift of solid state gas sensors. Submitted (1999). 10) Holmberg, M., Davide, FAM, Natale, C.Di., D'Amico, A., Winquist, F., Lundström, I .: Drift counteraction in odor recognition applications: Lifelong 10 15 20 25 30 35 40 ø u - o nu 518174 26 calibration method. Sensors & Actuators B 42 (No. 3) (1997) 185-194. 11) Jackson, J.E .: Principal components and factor analysis: Part 1. Principal Components. J. Qual. Tech. 13 (1) (1981). 12) Konstantinov, K., Chuppa, S., Sajan, E., Tsai, Y., Yoon, S., Golini. F .: Real-time biomass concentration monitoring in animal cell cultures. TIBTECH 12 (1994) 324-333. 13) Mandenius, C.F., Eklöv, T., Lundström, I. Sensor fusion with on-line gas emission multisensor arrays and standard process measuring devices in a Baker's yeast manufacturing process. Biotechnol. Bioeng. 55 (1997), 427-438.

An algorithm for least-squares Indust. 14) Marquardt, D.W .: estimation of non-linear parameters. J. Soc.

Appl. Math. 2 (No. 2) (1963) 431-441.

Schulze, B., Freitag, R., Scheper, On-line immunoanalysis 15) Middendorf, C., Th., Howaldt, M., Hoffmann, H .: for bioprocess control. J. Biotechnol. 31 (1993) 395-403. 16) Omstead, DR, Frame, KK, Tsao, EI, Price, GP, Glass, JC, Bohn, MA, McNamee, KP, Reese, JA, Munster, MJ, Jain, D .: Monitoring and control of animal cell cultures . IFAC Modeling and Control of Biotechnical Processes, Colorado, USA (1992) pp. 21-27. 17) Rumelhart, D.E., Hinton, G.E., Williams, R.J .: Learning internal representations by error propagation.

In: D.E. Rumelhart and J. McClelland (eds), Parallel data processing. M.I.T. Press: Cambridge, MA (1986). 27 TABLES Table 1. Overview of infection studies performed Response O _ P 2 hes elek- sensor- - Cultivation Infectious organism respgns åšïglsk [h after [h after infection] infection] 2 l rFVIII [0, l cfu / ml] Pseudomonas 21 15 CHO aeruginosa 2 l rFVIII [1 cfu / ml] Bacillus 10 1 CHO cereus 2 1 [100 cfu / ml] Pseudomonas _ reference- 14 1 aeruginosa medium 2 1 "[0.5 cfu / ml] Pseudomonas reference- 12 5 aeruginosa medium 2 l [4000 cfu / ml] Bacillus reference- _ 7 O subtilis medium 2 1 [20 cfu / ml] Bacillus reference- 13 6 subtilis medium 2 l [0.1 cfu / ml] Micrococcus reference- 39 10 luteus medium 5 28 Table 2. Calibration values for selected chemical sensors Sensor type Cultivation MF1 MF2 MF3 MF8 I 21.84 18.09 47.8 60.11 II 21.77 17.5 47.6 56.89 III 21.71 16.9 47.4 53 .68 IV 21.65 16.31 47.2 50.46 V 21.59 15.71 47 47.25 VI 21.53 15.11 46.8 44.03 Total operation 2% 16% 2% 27 5 Trendline adapted sensor responses on a bioreactor background under start-up conditions. ether.

Claims (12)

10 15 20 25 30 35 518174 29 PATENT REQUIREMENTS Method for the early detection of unwanted microbial infection in a cell culture using an electronic nose with a set of gas sensors including metal oxide semiconductor based field effect transistor (MOSFET) sensors, semiconductor metal oxide (MOS) sensors and CO2 sensors and capability pattern recognition, the method comprising the steps of taking a sample of released gas containing volatile compounds from the cell culture in a bioreactor, transporting the sample to the gas sensors of the electronic nose and comparing any specific gas sensor signal patterns with standard signal patterns corresponding to different microorganisms. of emitted gas is taken from an animal cell culture. A method according to claim 1, wherein the presence of bacteria, preferably Bacillus cereus and / or Pseudomonas aeruginosa, is detected. _ A method according to claim 1 or 2, wherein the cell culture medium is a recombinant CHO cell culture (kinetic hamster ovary cell culture) for the production of human blood coagulation factor VIII. A method according to any one of the preceding claims, characterized in that the electronic thereof, the nose gas sensor set, comprises 10 MOSFET sensors, 12 or 19 MOS sensors and a CO 2 sensor. A process according to any one of the preceding claims, wherein a sample of exhaust gas is taken from the bioreactor for about 30 seconds and about every 20 minutes. A method for estimating the product concentration in a cell culture by means of an electronic nose as defined in claims 1 and 4, comprising the steps of taking a sample of released gas containing volatile compounds from the cell culture in a bioreactor, transporting the cell sample to the electronic sensors of the electronic nose and calculating the product concentration by means of an artificial neural network. The method of claim 6, wherein the concentration of proteins, preferably recombinant human blood coagulation factor VIII in a CHO cell line culture, is estimated. The method of claim 6 or 7, wherein the cell culture is an animal or microorganism cell culture. A method according to claim 1, wherein the detection of any microbial infection is determined about 1-10 hours, preferably about Slh, previously for each microorganism compared to known techniques. A method according to any one of the preceding claims, wherein the electronic nose also includes quartz crystal based sensors and / or conductive polymer sensors and / or other semiconductor based sensors. 11. ll. A method according to any one of the preceding claims, wherein the methods according to claims 1 and 6 are carried out simultaneously in the same cell culture process. 12. Use of an electronic nose such as'defini-5. ' use in claim 1 for the detection of microbial infection, preferably bacterial infection, preferably infection of Bacillus cereus and / or Pseudomonas aeruginosa, in an animal cell culture. the concentration in a cell culture, preferably the concentration of proteins, most preferably the concentration of recombinant human blood coagulation factor VIII in a CHO cell line culture.