Real‐time dissolved carbon dioxide monitoring I: Application of a novel in situ sensor for CO2 monitoring and control

Abstract Dissolved carbon dioxide (dCO2) is a well‐known critical parameter in bioprocesses due to its significant impact on cell metabolism and on product quality attributes. Processes run at small‐scale faces many challenges due to limited options for modular sensors for online monitoring and control. Traditional sensors are bulky, costly, and invasive in nature and do not fit in small‐scale systems. In this study, we present the implementation of a novel, rate‐based technique for real‐time monitoring of dCO2 in bioprocesses. A silicone sampling probe that allows the diffusion of CO2 through its wall was inserted inside a shake flask/bioreactor and then flushed with air to remove the CO2 that had diffused into the probe from the culture broth (sensor was calibrated using air as zero‐point calibration). The gas inside the probe was then allowed to recirculate through gas‐impermeable tubing to a CO2 monitor. We have shown that by measuring the initial diffusion rate of CO2 into the sampling probe we were able to determine the partial pressure of the dCO2 in the culture. This technique can be readily automated, and measurements can be made in minutes. Demonstration experiments conducted with baker's yeast and Yarrowia lipolytica yeast cells in both shake flasks and mini bioreactors showed that it can monitor dCO2 in real‐time. Using the proposed sensor, we successfully implemented a dCO2‐based control scheme, which resulted in significant improvement in process performance.

reactions. For example, phosphofructokinase, which is a rate-limiting enzyme in the glycolytic pathway, is inhibited by low intracellular pH.
A dCO 2 level above 20% is reported to be a growth inhibitory factor for certain microbial as well as mammalian cells which necessitates an effective aeration strategy during scale-up of bioprocesses (Blombach & Takors, 2015). However, excessive stripping of dCO 2 is also detrimental to cell growth, which suggests that there is likely an optimal level of dCO 2 for cell culture (Mostafa & Gu, 2003). The organism used for production, the product of interest and the required quality of the product will certainly determine this optimal dCO 2 value. The concentration of dissolved oxygen (DO) and carbon dioxide vary over time due to respiration of the cells and hence these parameters can give the true metabolic signature of the culture broth if monitored in real-time. In addition, as an intrinsic property, partial CO 2 pressure of shake flask, benchtop lab-scale bioreactor, and the production bioreactor will differ significantly as it is an inherent consequence of high absolute pressure and mixing conditions. Hence, dCO 2 concentrations must be analyzed carefully and thoroughly during scale-up and technology transfer activities (Matsunaga, Kano, Maki, & Dobashi, 2009;Mitchell-Logean & Murhammer, 1997;Mostafa & Gu, 2003). The real-time dCO 2 values will be useful not only for optimizing but also for reproducing culture conditions independent of scale (Ahuja, Shilpa Jain, & Ram, 2015). Thus, the careful monitoring and control of DO and dCO 2 are needed in assessing the culture conditions (Ge, Kostov, & Rao, 2003;Gupta & Rao, 2003).
With Food and Drug Administration's process analytical technology (PAT) drive and rigorous quality requirements especially in biologics manufacturing (Food and Drug Administration, 2004), the attention towards continuous process supervision and real-time control is greater than ever Gomes, Chopda, & Rathore, 2018).
Despite this fact, small-scale processes are still lagging in terms of generating real-time data due to a lack of suitable online sensors.
Traditional electrochemical sensors are usually not used because they are bulky and invasive. Disposable optical sensors are small and only partially invasive, but there are concerns regarding the toxicity of the patch and the phototoxicity of the illuminating light (Ge & Rao, 2012;Ge, Kostov, & Rao, 2005;Gupta & Rao, 2003). Over the past decade, new sensor technologies have become available to monitor the performances of shake flasks and small-scale fermentation systems (Table 1). Among these, the oxygen transfer rate (OTR) device and respiration activity monitoring system (RAMOS) measure OTR and respiration activities (not direct O 2 and CO 2 concentration) in the headspace of a shake flask and not the dissolved CO 2 .

BlueSens gas sensor GmbH
Liquid and gas Used RAMOS to measure in the headspace and rotating flexi-tube optical sensor to monitor dissolved O 2 , CO 2 , and culture pH Hansen, Jacob, Luchterhand, and Büchs (2012) Fluorescence-based method for monitoring CO 2 and O 2 in the culture broth and in the headspace of an Erlenmeyer flask and broth pH as well Ge and Rao (2012) Circulation direct monitoring and sampling system-a system that continuously extracts and measures liquid (by electrode) and gas-phase samples Takahashi et al. (2017) Takahashi and Aoyagi (2018a, 2018b, and 2018c Abbreviations: dCO 2 , dissolved carbon dioxide; OTR, oxygen transfer rate; RAMOS, respiration activity monitoring system. in the headspace of shake flask with a culture plug but not the dissolved CO 2 (BlueSens gas sensor GmbH). In all these three systems sensors or a measuring device rotates with the shake flask which may sometime cause drifts in the signal. The fluorescencebased method has been reported to measure O 2 and CO 2 in the headspace as well as in culture broth. The effect of fluorescent pigment has always been debated (Ge et al., 2005;Gupta & Rao, 2003 (Takahashi & Aoyagi, 2018a, 2018bTakahashi, Sawada, & Aoyagi, 2017). However, it is critical to prevent cells from clumping to avoid the clogging of the liquid circulation part.
In this article, we describe a novel method that correlates the concentration of dissolved CO 2 in culture broth with the initial diffusion rate of the CO 2 across a gas-permeable tubing. Silicone tubing was fully immersed in the culture medium through which CO 2 gas diffuses which was then circulated to a CO 2 sensor equipped with a custom-designed microcontroller pump and valves, located outside of the cell culture vessel. We have decided to select two model organisms for this study having distinct respiration activities.
Baker's yeast was selected because it is well-known for its Crabtree effect. Through its Crabtree or overflow metabolism, yeast cells can switch from purely oxidative metabolism to a respiro-fermentative The implementation of our novel sensor in monitoring dCO 2 in shake flask and in mini bioreactors was demonstrated using yeast fermentation as a case study. We were able to demonstrate the distinct dissolved CO 2 profile for these selected two organisms and further demonstrated a control strategy to control the metabolism.
We believe that the enabling PAT at such small scale of operation will allow users to develop a scaled-down model of pilot-scale bioreactor operations. This will not only make the process development costeffective and faster but also expected to follow a smoother transition of scale-up and technology transfer activities.

| Dissolved CO 2 sensor
This novel method correlates dCO 2 concentration in cell culture with the initial diffusion rate of the CO 2 across gas-permeable silicone tubing, which was provided by Sartorius (Goettingen, Germany) and originally used for aeration in bioreactors (BB-8848017). Silicone tubing (ranging from 4 to 6 cm in length) allows for the diffusion of CO 2 across its wall when fully immersed in the culture medium and connected to a CO 2 sensor (K30 CO 2 sensor from CO2Meter.com).
The sensor is a fast-response infrared CO 2 sensor, which has a measurement range of 0-10,000 ppm with a precision of 20 ppm. The sensor is designed to output digitized values in serial format (RS 232).
It is connected to a computer via serial-to-USB converter (i.e., FT232 from Future technology). The system is equipped with two 3-way valves and a pump, which are located outside of the cell culture vessel. The gas pump is a microdiaphragm pump produced by Parker (Hollis, NH), which provides a flow rate of up to 800 ml/min. The 3-way valves are solenoid valves produced by the Lee Company (Westbrook, CT). They require only a very small power to operate (~3 mWh/actuation), have small dimensions (<2″ in any direction).
Versilon F-5500-A tubing, which is reported to have significantly low CO 2 diffusivity, was used to connect the system components together (shown as black lines in Figures 1-3). The electronics for activating the valves and digitizing the readings of the sensors are controlled by a dedicated microcontroller, which is also responsible for communicating the data to the computer (Ge et al., 2018). The program is written in LabVIEW, which controls the valves, reads the sensors, logs the data, and calculates the slopes in real-time. All the time intervals are user adjustable so as to allow flexibility of the software.
Before a measurement cycle, ambient air outside the culture vessel was pumped into the system (Air IN) and allowed to flow out through a separate outlet (Air OUT). This process purges any CO 2 in the tubing system and returns it to normal atmospheric levels (~400 ppm). After the purging process is complete, the valves close, allowing CO 2 from the culture to diffuse across the gas-permeable F I G U R E 1 Illustration of the operational working of the dissolved carbon dioxide sensor. In the purge mode, the ambient air outside the culture vessel was pumped into the system (Air IN) and allowed to flow out through a separate outlet (Air OUT). This process purges any CO 2 in the tubing system and returns it to normal atmospheric levels (~400 ppm). On the other hand, in recirculation mode, the Air IN valve closes, allowing CO 2 from the culture to diffuse across the gas-permeable tubing into the system. The pump recirculates the diffused gas through the CO 2 sensor tubing into the system. The pump recirculates the diffused gas through the CO 2 sensor. The purge and recirculate modes of operation and sensor setup are illustrated in Figure 1. The measurement principle and the operation procedure have been described in detail previously (Chatterjee et al., 2015). For a shake flask, the silicone tubing was connected to the gasimpermeable Versilon tubing with simple luer fittings. A spring was inserted through the tube to avoid blockages, bents or kinks. For shake flask, aseptic sampling was carried out through a Versilon tube inserted in it so as to have minimal disturbance in O 2 and dCO 2 balance, which might change significantly if sampling carried out by conventional way of opening the culture plug (Takahashi and Aoyagi, 2018c). A cylindrical mini bioreactor (2.5 cm inner diameter) of 100 ml capacity was used for controlled feeding experiments. For the mini bioreactor, a probe construct was used that is similar in design to typical PAT probes used in larger bioreactors. Figure 3 outlines the direction of airflow through the top of the probe, down through a needle that is surrounded by silicone tubing, back up through the needle housing, and out the side of the probe. This allows for diffusion rate measurements to be collected in a narrow culture vessel with a low volume.

| Sensor characterization
After the prototype was built, the characteristics of the sensor such as sensitivity, the limit of detection (LOD), and so forth were studied. To avoid the accumulation of CO 2 in the system, the purge stage should last long enough. After some initial experiments, the following durations were determined to be the most appropriate:

| Strains and culture conditions
The developed sensor was used to monitor dCO 2 in both baker's yeast and Y. lipolytica P01g-Leu fermentations. DO was monitored by an optical sensor that is provided by Scientific Industries, Inc. The detailed description of the DO optical sensor can be found in the literature (Tolosa, Kostov, Harms, & Rao, 2002). The cultures were sampled once a day for optical density measurement.

| Baker's yeast fermentation
Fleischmann's Active Dry Yeast (Saccharomyces cerevisiae yeast, also known as baker's yeast) was used in these experiments and was directly added to yeast extract-peptone-dextrose (YPD) media incubated in shaker flasks and mini bioreactors. The media was prepared using individual components obtained from Sigma-Aldrich. For sensor characterization (proof-of-concept) run, the concentrations were as: 10 g/L of yeast extract, 20 g/L of each peptone and dextrose. For the control experiments, the dextrose concentration was reduced to 3 g/L deliberately while other components were at the same concentrations.
The lower initial dextrose concentration gave us the flexibility to test the CO 2 control strategy by adding additional glucose. The initial yeast dry mass and the YPD media concentration were noted for each experiment involving this strain. Dry yeast (0.2 g) was added to 50 ml of YPD media in a shake flask and allowed to grow at 30°C at 250 rpm for 4 hr. This grown inoculum was further used to start the experiments with the desired initial inoculum density.

| Y. lipolytica P01g-Leu fermentation
Y. lipolytica is classified as an oleaginous yeast species because of its ability to accumulate lipids in large quantities (Xu, Qiao, & Stephanopoulos, 2017). We used a genetically modified version of a leucine auxotroph strain (P01g-Leu) that requires a leucine complementation pathway to grow (Xu, Qiao, Ahn, & Stephanopoulos, 2016). This strain was engineered for flavonoid biosynthesis (P01g with flavonoid pathway). The P01g-Leu strain was cultured in the YPD media because of its inability to synthesize leucine (leucine is needed in the culture media to grow). Three hundred microliters of glycerol stock were added to 5 ml of YPD media in 50 ml Falcon tube and allowed to grow at 30°C at 250 rpm for 20-24 hr. This preculture was further used to inoculate the shake flask culture with the desired starting optical density. The slope of the lines represents the sensitivity of the sensor.

| Sensor characterization
According to our previous studies (Chatterjee et al., 2015), the sensitivity of the rate-based sensor is affected by the mass transfer coefficient of CO 2 in the sampling tube material (k), the total mass transfer area of the sampling tube (A), and the internal volume of the sensor system (V): Although the K30 CO 2 sensor has much lower precision than the LI-820 CO 2 gas analyzer (20 vs. 1 ppm), the K30-based dCO 2 sensor has a much higher sensitivity than the LI-820-based dCO 2 sensor (~3.5 vs. 0.9 ppm/%) due to its much lower volume. Increasing the mass transfer area (the size) of the sampling tube can increase the amount of CO 2 diffusing into the system. As a result, a higher signal will be detected at the same dCO 2 level. Thus, the sensitivity of the system will be increased.
Using a more permeable material for the sampling tube, the sensitivity of the system can be increased. However, using a too permeable material could saturate the CO 2 sensor, resulting in erroneous readings.
The LOD of the sensor, which is the lowest dCO 2 that the sensor can reliably measure is defined as three times of baseline signal noise divided by the sensitivity. This value was determined by conducting multiple baseline diffusion rate measurements and calculating the standard deviation for the dataset. Because LOD is inversely proportional to the sensitivity, any measures discussed above that can increase the sensitivity of the sensor can improve its LOD. In addition to modifying the parameters shown in Equation (1), the lengths of time for each stage of the measurement cycle could also affect the LOD. The length of the purge stage is important because it needs to be long enough to fully remove the CO 2 buildup from the system before the measurement stage begins. Increasing the duration of the recirculation stage can increase the repeatability (decrease the noise) of the measurement, so it can improve the LOD.
However, increasing the recirculation duration will make the measurement cycles longer. In the tests described in this paper, the flush stage lasted 30 s and the recirculation stage lasted 60 s. So each measurement cycle lasted 1.5 min. For most applications, 1.5-min measurement intervals are fast enough, but for applications where accuracy is the highest priority, the duration of the recirculation can be increased for better accuracy. At the above conditions, the LOD was calculated to be 0.04% for Bravo-01, 0.13% for Bravo-04%, and 0.05% for K30. As mammalian cell cultures are usually maintained at 5% CO 2 , the dCO 2 in bacteria and yeast fermentations can be higher due to faster growth rate, the described dCO 2 sensor is sensitive enough for monitoring dCO 2 concentration in fermentations.

| Shake flask monitoring
Experiments were performed using baker's yeast and Y. lipolytica P01g-Leu strain at shake flask scale in batch mode using YPD media.
Process data for these preliminary experiments were successfully  where the significant distinction in their metabolic pattern can be seen. The initial sharp rise in dCO 2 concentration (first 7 hr) in the case of baker's yeast was probably due to the Crabtree effect where respire-fermentative glucose metabolism led to the production of CO 2 and ethanol Persad, Chopda, Rathore, & Gomes, 2013).
Thereafter, dCO 2 concentration decreased indicating exhaustion of glucose. From 9 to 20 hr, there was around 2% increase in dCO 2 concentration which indicates that the produced ethanol was being consumed (Figure 5a). In contrast, the yeast Y. lipolytica P01g-Leu followed a respiratory metabolism and therefore, dCO 2 rose steadily with a concomitant drop in DO concentration. The overall dCO 2 concentration was higher in yeast Y. lipolytica P01g-Leu compared with baker's yeast. As shown in Figure 5c, the fermentation period after 48 hr showed a higher OD in baker's yeast, which was around 3.5 even after starting with a lesser initial OD (the first F I G U R E 4 Calibration curves for three different prototypes showing the initial diffusion rate of CO 2 through the sampling tube versus dissolved carbon dioxide concentration. K30 was built with a K30 CO 2 sensor. Bravo-01 and Bravo-04 were built with a LI-820 CO 2 gas analyzer. The calibrations were performed in triplicate. The standard deviations are given as error bars, which are invisible because they are smaller than the symbols [Color figure can be viewed at wileyonlinelibrary.com] data point was somewhat fluctuating due to the small number and probably sampling error), compared with the OD in Y. lipolytica P01g-Leu yeast, which was around 3.3 even after starting with a higher initial OD of 0.5. This may suggest a higher CO 2 concentration (~12%) in the case of Y. lipolytica P01 Leu yeast might have inhibited cell growth compared with the baker's yeast fermentation where dCO 2 concentration reached only 5%. This is in accordance with various literature data that have shown a similar inhibitory impact of higher CO 2 levels (Blombach & Takors, 2015). In addition, a similar distinct dCO 2 pattern in the Crabtree positive and negative strains was observed when measured in headspace by the RAMOS system (Anderlei et al., 2004).

| dCO 2 control by controlled glucose feeding
The concentration of glucose is critical in the baker's yeast culture as excessive glucose may turn on the Crabtree effect even in oxidative conditions. As glucose metabolism yields to biomass, CO 2 and water, we gave the target setpoint for the dCO 2 by keeping glucose in feedback control (on-off control). In the first condition, we controlled the dCO 2 levels at 10% for 4-8 hr using 400 g/L glucose and then turned off the controller for the remaining process duration. In other conditions, we controlled the dCO 2 levels at 10% and 6% for 4-24 hr using 400 and 200 g/L glucose feed concentration, respectively. The choice of maintaining a set of dCO 2 values using a certain concentration of glucose solution has been taken from our preliminary experimentation (data not included) while testing the sensor.
From Figure 6a-c we can see that the feedback control was able to maintain the dCO 2 levels around the set target. The experiment in which feeding was performed for only 4-8 hr showed that dCO 2 levels oscillated at around 10% and thereafter declined as the feeding was stopped (Figure 6a). This condition resulted in lower biomass concentration and reached only OD around 8 (Figure 6a).
The experiment in which dCO 2 maintained at around 10% (4-24 hr), DO remain relatively at lower levels (5-10%) as shown in Figure 6b and OD reached to 14.5. In the third case, we reduced the concentration of the feed solution and set the dCO 2 levels to around 6%. With controlled glucose addition, dCO 2 remained well controlled at the desired set point of 6%, which led to significantly higher biomass (OD reached 25.2) as shown in Figure 6c). The biomass yield over glucose in the case of 400 g/L glucose fed (for 4-24 hr) experiments was estimated to be 0.47 OD/g-glucose while in the case of 200 g/L glucose fed (for 4-24 hr), it was estimated to be 1.0 OD/g-glucose. With a 4% lower target dCO 2 concentration, the biomass yield was increased by 114% (Figure 6a-c). This can be explained by the fact that with the condition in which 10% dCO 2 was maintained, the more glucose pumped into the bioreactor in an oxidative condition might have led to more ethanol formation (Crabtree effect) as a byproduct, which resulted in low biomass concentration (OD = 14.5). In contrast, the other condition in which dCO 2 was maintained to only 6% was able to maintain the culture in the respiratory regime and resulted in more biomass. Thus, a For example, in Figure 5, it is hard to predict the existence of the Crabtree metabolism of the organism solely from DO response. In the growth phase, the DO signal went to 0 value and when the cells entered the stationary phase, the DO signal rose to 100% ( Figure 5).
On the other hand, dCO 2 provides a more nuanced response that can be correlated to biomass production. This is due to CO 2 being the final product of respiration both in microbial as well as in mammalian processes. It has a strong impact on cell metabolism in multiple ways as discussed in Section 1 of this paper. Thus, the real-time dCO 2 profile can give the true metabolic signature of a process. However, until now, CO 2 is mostly monitored in the exhaust gas, which itself cannot give real-time culture broth conditions as the information captured is filtered through the headspace. This means there exists an inherent lag in relaying the dCO 2 concentrations to the gaseous phase CO 2 . Very few sensors are available in the market which can measure dCO 2 concentration, and most will not fit in small-scale systems such as shake flasks or mini bioreactors.
Here we have presented the application of a novel, noninvasive,  The lowered concentration of dCO 2 resulted in a significant increase in biomass productivity ( Figure 6). The experiment performed for sensor characterization ( Figure 5) is the run in which no dCO 2 control was implemented that resulted in a maximum OD of around 3.5 at around 24 hr. On the other hand, all the dCO 2 controlled experiments ( Figure 6) yielded a higher OD. The highest OD of 25.2 was achieved when dCO 2 was controlled at 6%, which is almost a seven-time increase in OD compared with uncontrolled conditions.
Many researchers have used off-gas measurements to estimate the biomass in the reactor. The substrate demand is then determined according to the amount of biomass. However, this method is still an "open loop" control strategy and may lead to a lag in control action (Muthuswamy & Srinivasan, 2003;Valentinotti et al., 2003). Other case studies in the benchtop, as well as large scale bioreactors, emphasize the important role of dCO 2 monitoring and its significant impact on cell growth and protein production (Mitchell-Logean & Murhammer, 1997; Mostafa & Gu, 2003).
A similar phenomenon occurring in shake flask culture, where the performance of the shake flask in terms of biomass growth and product concentration was significantly affected by dCO 2 concentrations, was reported (Takahashi & Aoyagi, 2018a, 2018b. In one of the more interesting findings, researchers have shown that opening the shake flask closure for sampling allows diffusion of gases into the atmosphere, such that CO 2 concentrations decrease temporarily in the broth as well as in the headspace, thereby significantly impacting the community structure of soil microbes (Takahashi & Aoyagi, 2018b). A study conducted by McIntyre and McNeil (1997) concluded that culture is more vulnerable to CO 2 inhibition in the lag phase. In addition, during the inoculation step, the culture can experience many pressures due to the significant differences in the environment of a shake flask and in the bioreactor conditions.
Monitoring dCO 2 levels across all scales of biomanufacturing will enable control of the culture environment to prevent any severe shock to the growing cells.
We believe that the studies presented in this paper using our in situ dCO 2 sensor proves the criticality and importance of monitoring and control of dCO 2 . Monitoring dCO 2 will not only indicate real-time culture conditions but also give the critical process information for taking effective control decisions should deviations occur. We also showed that the proposed sensor is flexible in configuration and can be easily fit in shake flasks and other small-scale mini bioreactors. This will enable us to extract more process information even from small-scale systems thereby accelerating process development many folds.

| CONCLUSION
Shake flask and the small-scale fermentation processes are generally hindered by the scarcity of in-line sensors. We have identified that there is a need to develop a portable dCO 2 sensor for such smallscale fermentation due to the significant effect of dCO 2 on the overall metabolism and the subsequent scale-up activities. In this manuscript, we present a novel, rate-based technique for monitoring dCO 2 in cell cultures by measuring the initial diffusion rate across a silicone sampling tube immersed in the culture media. Demonstration experiments conducted with baker's yeast and Y. lipolytica yeast cells in both shake flasks and mini bioreactors show that it can monitor dCO 2 in real-time. Using the proposed sensor, we successfully implemented a dCO 2 -based control scheme, which resulted in significant improvement in process performance. Through the implementation of dCO 2 based control strategy, we have shown that by controlling dCO 2 at various levels we can direct the metabolic fate of the process in real-time. In the future, our efforts will be directed towards using the developed dCO 2 based control scheme to reduce the disparity that exits across the different scales of operation.

ACKNOWLEDGMENT
The present work was conducted under the sponsorship and generous support of the Bill and Melinda Gates Foundation.