Optimization of medium with perfusion microbioreactors for high density CHO cell cultures at very low renewal rate aided by design of experiments

A novel approach of design of experiment (DoE) is developed for the optimization of key substrates of the culture medium, amino acids, and sugars, by utilizing perfusion microbioreactors with 2 mL working volume, operated in high cell density continuous mode, to explore the design space. A mixture DoE based on a simplex‐centroid is proposed to test multiple medium blends in parallel perfusion runs, where the amino acids concentrations are selected based on the culture behavior in presence of different amino acid mixtures, and using targeted specific consumption rates. An optimized medium is identified with models predicting the culture parameters and product quality attributes (G0 and G1 level N‐glycans) as a function of the medium composition. It is then validated in runs performed in perfusion microbioreactor in comparison with stirred‐tank bioreactors equipped with alternating tangential flow filtration (ATF) or with tangential flow filtration (TFF) for cell separation, showing overall a similar process performance and N‐glycosylation profile of the produced antibody. These results demonstrate that the present development strategy generates a perfusion medium with optimized performance for stable Chinese hamster ovary (CHO) cell cultures operated with very high cell densities of 60 × 106 and 120 × 106 cells/mL and a low cell‐specific perfusion rate of 17 pL/cell/day, which is among the lowest reported and is in line with the framework recently published by the industry.


| INTRODUCTION
Monoclonal antibodies (mAbs) for therapeutic applications are of paramount importance for the biopharmaceutical industry as mAbs and their variants represent approximately half of all newly approved products (O'Flaherty et al., 2020). Manufacturers of mAbs are, however, challenged with the production to meet the increasing market demand. Recent advances in cell line and process development have led to substantial improvements in the productivity of cell culture processes (Shukla et al., 2017;Tihanyi & Nyitray, 2020).
Intensification realized thanks to high cell density perfusion processes and continuous downstream processing are recognized as promising avenues. A prominent role is attributed to the development of culture medium as it has a great impact on the cell and process performance. Due to the high complexity of media for mammalian cell culture, one-factor-at-a-time optimization strategy has been found to be ineffective and instead the statistical approach of design of experiments (DoE) is most often adopted to reduce the number of experiments (Ritacco et al., 2018). DoE has been applied to improve the production of mAbs in Chinese hamster ovary (CHO) cell lines by analyzing the effect of components with a response surface methodology and identifying their optimum levels (Liu et al., 2015;Torkashvand et al., 2015). Another experimental approach, namely metabolic profiling, has demonstrated its effectiveness in optimizing mAb production through the design of feed regimes relieving metabolic bottlenecks from nutrient limitations (Kishishita et al., 2015;Sellick et al., 2011).
The majority of proposed medium development strategies were established for cells grown in batch and fed-batch culture, until now the most dominant cultivation mode for mammalian cells. The realization of process intensification, generally by increasing the cell density and cultivation time in a continuous system, has sparked great interest by the industry to apply perfusion processes for mAb manufacturing (Chen et al., 2018). In a steady-state perfusion bioreactor, the cell concentration is maintained at a constant level through cell removal at a rate corresponding to the growth rate, that is, operating a cell bleed, while the culture medium is continuously renewed, potentially generating cell densities multiple times higher compared to fed-batch cultures (Clincke et al., 2013). Besides the increase in volumetric productivity, the steady-state condition can add the benefit of obtaining product with consistent quality (Gomis-Fons et al., 2020).
With the advent of continuous processing for mAb production, the objectives of the process development are different from fedbatch process and traditional medium design strategies can become obsolete. A key parameter in perfusion, namely the cell-specific perfusion rate (CSPR) determining the ratio of medium renewal rate per cell density, is often minimized to reduce the medium consumption, while maintaining the process performance at the desired level (Schwarz, Gomis-Fons, et al., 2022;Schwarz, Mäkinen, et al., 2022). A procedure for the design of a perfusion process can involve finding the minimum CSPR for a given medium and subsequently increasing the cell density at constant CSPR. The operational boundaries such as limitations in gas transfer of the bioreactor used for the commercial production, dictating the potential supported cell density, should also be taken into account. In addition, tuning the medium composition is an essential part, in particular for operation at low CSPR to support high viability and productivity over extended process times. Adjusting the concentration of components, such as the sugars, provides also the possibility to tune the N-glycosylation pattern of proteins .
A difficulty in performing extended medium screenings and optimizations in perfusion is the lack of scale-down models with equivalent performance to perfusion stirred-tank bioreactors (STBRs). The complexity of a perfusion bioreactor with a cell retention system and feed, harvest and cell bleed lines, is challenging to obtain in a parallel format in milliliter-scale systems. Although not being a continuous system, the spin tube operated in semi-perfusion mode is a model that has been extensively used to mimic perfusion bioreactors (Gomez et al., 2017;Mayrhofer et al., 2021;Villiger-Oberbek et al., 2015). Such a model can be a supportive tool to assist process development with resource-intensive perfusion bioreactors.
Automated high-throughput miniaturized bioreactors with integrated sensors, such as ambr ® 15, provide a better control of process parameters, and have taken a firmly established role in process development. Such systems have also been applied to mimic perfusion mode in a batch-refeed manner (Jin et al., 2021). Recently, a microbioreactor (MBR) system in perfusion mode with integrated cell retention system has been introduced, the Erbi Breez™. This automated microfluidic bioreactor with 2 mL working volume could be a valuable tool to support process development, and accelerate its timelines, replacing or complementing lab-scale perfusion STBR experiments.
In the present study, we developed a medium optimization strategy applicable to high cell density perfusion cultures of mAb producing CHO cells for low CSPR. According to the recently proposed framework for integrated and continuous biomanufacturing (Coffman et al., 2021), the medium should support cell densities of up to 120 × 10 6 cells/mL with a maximum perfusion rate of two bioreactor volumes/day (vvd), which translates to a CSPR of 17 pL/cell/ day or lower. We focused on the optimization of key components in the medium, namely selected amino acids and carbohydrates, in medium used for high cell density perfusion process at a CSPR of 17 pL/cell/day. These substrates are present in the largest quantities in the medium and can be quantified by standard analytical techniques, alleviating their tuning in the medium. Our strategy involved a three-step procedure with (1) initial screening in semiperfusion mode, (2) design of medium formulations for optimization in a simplex-centroid mixture design experiment utilizing microfluidic perfusion MBRs, and (3) validation of the optimal medium composition in STBRs. In addition, we provided a head-to-head comparison of the perfusion culture performances under similar conditions in the different scale-down models.

| Cell line and culture medium
Experiments were performed with a CHO-K1 cell line with glutamine synthetase (GS)-based expression system producing a mAb (trastuzumab). After cell thaw, the cells were expanded in HyClone ActiPro medium (Cytiva) for three passages, that is, lasting a total of 7 days, in shake flasks at 37°C, 5% CO 2 atmosphere, and 120 rpm shaking speed before inoculation into bioreactors. For perfusion cultures, a prototype medium, named "N0," based on the ActiPro formulation with reduced concentration of amino acids and without glucose was used. N0 medium was then supplemented with amino acid and sugar stock solutions to adjust the nutrient concentrations, generating new media called M1-M4, according to the experimental design given in the text. The pH was set in all media to 7.4 and the osmolality was measured by an Osmomat 3000 (Gonotec) to confirm an acceptable range of 345-365 mOsm/kg.

| Design of experiments (DoE) for medium optimization
DoE was supported by Modde software (Sartorius Stedim). A simplex-centroid mixture design was selected with four media, resulting in different mixtures originating from N0. An advantage of simplex-centroid mixture design is that for each condition the media are mixed in proportions between 0% and 100% (or fractions between 0 and 1) such that all the fractions of the different media always add up to 1 in one condition, decreasing the risk for precipitation or excessive osmolality. The conditions were tested in steady-state perfusion cultures, where each condition was maintained for a period of 7-8 days to ensure that the culture reaches a steady state. The perfusion cultures were operated at a constant viable cell density (VCD) of 55-60 × 10 6 cells/mL maintained by cell bleeding, which was reached on average 2 days after inoculation at a seed VCD of around 30 × 10 6 cells/mL from shake flask cultures concentrated by centrifugation (200g, 5 min). The perfusion rate was kept constant at 1 reactor vvd throughout the cultures. Nine perfusion cultures were conducted in total, for most of them with two conditions per culture, that is, from Day 3 to 10, followed by a medium change on Day 10 generating a new condition from Day 10 to 17. The last 3 days of each condition was considered at steady state and taken into account to study the effect of the different conditions. For each condition, the data of these last 3 days were considered as replicate.

| Perfusion MBR system
The conditions proposed by the DoE were conducted in an automated perfusion system with 2 mL working volume. Four parallel Erbi Breez MBRs (Erbi Biosystems) were employed to perform the perfusion cultures in several experimental runs. Validation experiments with optimum medium composition for steady-state perfusion as determined by the DoE study were conducted in further parallel MBR runs as duplicates.
The technical functionality of the MBR system shown in Figure 1 is briefly described here. As shown in the simplified side view in Figure 1a, two plastic layers separated by a thin silicone membrane define the channels and valves. The silicone membrane is actuated with pressure or vacuum from the upper air side to deflect the silicone membrane and seal off or open channels. The actuation structures are designed with different geometries and sizes to create valves, reservoirs, and culture chambers. Three deep 1 mL volume actuation sections, that is, chambers, are connected in a ring and form the growth chamber, that is, where the cell cultivation is carried out.
Only two of the 1 mL chambers are shown in the side view of Figure 1a for simplicity. During operation, two of the chambers are filled and one chamber is kept empty resulting in a nominal total working volume of 2 mL. Inflation of the silicone membranes in the growth chamber sections pushes the culture through wide channels between the three chambers to provide mixing. In addition, because the silicone membrane has high gas permeability, there is efficient air, oxygen, and carbon dioxide gas exchange with the culture fluid via gas diffusion, allowing the reactor to maintain adequate oxygenation at high cell density without bubbles. Thanks to the bubble-free oxygenation and mixing, optical density (OD) measurements can be carried out directly through the chamber to estimate online the total cell density. Closed-loop control of the cell density is performed based on this online monitoring. There are two OD sensors, one sensor (OD2) used as default, and an extra sensor (OD1) with a shorter path length to increase the dynamic range if needed. The OD signal gives an information of total amount of viable and dead cells.
One optical dissolved oxygen (DO) and two pH sensors (allows for drift correction and redundancy) are also embedded in the base of the reactor for contact-free online measurements. The sensor data are communicated to a pneumatic controller which actively adjusts the oxygen and CO 2 concentration of the gas used to pressurize the growth chamber sections during mixing for feedback control of the DO and pH. An alkali solution can also be added from the input pump to control the pH. For cell culture experiments, the integrated sensors were used to monitor and control the temperature at 37°C, DO at 40%, and pH at 7.0. The permeable silicone membrane allowed for gas transfer to regulate the DO with O 2 and the pH with CO 2 . A minimum CO 2 concentration of 5% was selected in the inflow gas atmosphere.
Additionally, 0.5 M Na 2 CO 3 solution was used to control the pH. The addition of base and perfusion medium, as well as the removal of cell suspension for bleeding and sampling, was achieved with built-in microfluidic channels, valves, and precalibrated input and perfusion pumps, described above. The cell separation was performed

| Perfusion STBR
Validation experiments were performed in 200 mL STBR by perfusion through an alternating tangential flow filtration (ATF) or tangential flow filtration (TFF) system with external hollow fiber cartridges (model CFP-4-E-3MA, Cytiva) as previously described . For the TFF system, a PuraLev-i30SU pump (Levitronix) was used to recirculate the cell suspension through the hollow fiber.
The flow rate was controlled at 0.4 L/min with a Leviflow LFSC-D flow sensor (Levitronix). The same flow rate was set in the ATF2 system (Repligen). Setpoints for temperature, pH, DO, target VCD, and perfusion rate were set as in the MBR runs. The STBRs were inoculated at 2.5 × 10 6 cells/mL, and the culture was expanded over a course of 7 days to reach the target VCD and initialize continuous cell bleeding. Perfusion was initiated 1 day after inoculation at 0.5 vvd, 0.75 vvd on Day 4, and finally 1 vvd on Day 5. For the first three days of perfusion, standard ActiPro medium was used, and then switched to optimized "validation medium." A 5 mL sample was taken once a day for the subsequent offline analyses described in Section 2.6.

| Semi-perfusion cultures
Semi-perfusion cultures were performed in 50 mL TPP TubeSpin ® Bioreactors (Corning), referred here as spin tube BRs, according to a procedure similar to previously published work by other groups (Gomez et al., 2017;Mayrhofer et al., 2021;Villiger-Oberbek et al., 2015). In the present study, the target seed cell density was 18 or 50 × 10 6 cells/mL and the working volume 10 mL. The spin tube BR cultures were inoculated at target cell density. Every day, these cultures were sampled for cell count and viability (measured with technical duplicate), a part of the cell culture was discarded, and the rest was centrifuged (5 min at 200g). The supernatant was then removed and stored for spent media analysis. The cell pellet was resuspended in 10 mL fresh medium and a new sample was taken to measure the cell density, ensuring that the reseeded cell density was the targeted one. For the initial screening experiment, the target cell density of the spin tube BR was 18 × 10 6 cells/mL used for F I G U R E 2 Inoculation, sampling, and volume correction procedures. Only two of the three chambers are shown in the diagram and the third chamber (not shown) is assumed to be full for inoculation and volume correction. (1) Inoculum is introduced into the reactor through the sampling port (a), filling two of the three chambers to the standard fill volume.
(2) In standard operation, the chambers are nominally in the relaxed state. (3) During sampling, the chamber is first prediluted with medium to increase the volume of the reactor. (4) For volume correction and sampling (equalization), one of the chambers is pressurized, resulting in any extra volume of 100 μL being trapped under the pressurized silicone membrane. (5) To remove the extra volume, one of the three output ports, (a) sample, (b) waste, (c) harvest, is opened and the additional volume injected into the chamber is removed restoring the reactor to the standard fill volume. inoculation and reseeding, and the shaking speed was 320 rpm. For the validation experiment, the target cell density was 50 × 10 6 cells/ mL and the shaking speed was 350 rpm. The temperature and pCO 2 setpoints of the incubator were 37°C and 5%.

| Analytical methods
The viable/total cell density and viability were measured with a Norma XS (iPRASENSE). Supernatant from the spin tube BR and harvest collected during the past 24 h of STBR and MBR cultures were analyzed for the glucose, lactate, glutamine, ammonium, and mAb concentrations with a Cedex Bioanalyzer (Roche Diagnostics).
The mAb concentration was additionally analyzed in the supernatant from samples daily taken from the STBR or the MBR cultures after removing the cells by centrifugation. Bioreactor harvest and spin tube BR supernatant of selected samples was purified with protein A coated magnetic beads (MAGic Bioprocessing) for further analysis of mAb quality attributes, including N-glycosylation and charged variants. The methods are described in detail elsewhere (Brechmann et al., 2021). Amino acid analysis was performed with an AccQ Tag™ Ultra derivatization kit on an ACQUITY UPLC H-Class system equipped with an AccQ Tag Ultra C18 column, and tunable UV (260 nm) and fluorescence detector (λEx: 266 nm, λEm: 473 nm) (all Waters).

| Calculation of specific rates
The specific growth rate μ (day −1 ) and mAb productivity q p (pg/cell/ day) between 2 days were calculated as follows for the different operation modes: (2) X represents the VCD (1E6 cells/mL), t the time (days), P the mAb titer (mg/L), and subscripts 1 and 2 consecutive days.
Perfusion (at steady-state) in MBR and STBR: H represents the harvest rate in vvd (day −1 ) andḂ the bleed rate (day −1 ). The superscript BR stands for the titer in the bioreactor and HT for the titer in the harvest collected during 1 day. X̅ is the average cell density and P̅ BR 1,2 the average mAb titer over 2 days.
In the MBR, once the OD-based feedback control of the cell density was activated, it was assumed that X 2 was equal to X 1 so that the growth rate, μ MBR , could be approximated as: In practice, it was observed that Equation 5 gave estimations of the growth rate more accurate than Equation 3 since the fluctuations and measurement errors of the cell density were not taken into account in Equation 5.

| RESULTS
3.1 | Preliminary screening of the effect of amino acids in semi-perfusion culture-Screening-exp#1 As previously mentioned, an objective was to optimize key components of a perfusion medium, namely selected amino acids and carbohydrates, for high cell density perfusion process at CSPR 17 pL/cell/day. The work was based on medium ActiPro, for which a strategy to improve the glucose and amino acid contents was developed and applied. For the cell line used in this work, it had been possible to push down the CSPR to 29 pL/cell/day in high cell density perfusion process using ActiPro medium however lowering the CSPR under this value required reinforcement of the medium. A CSPR of 33 pL/cell/day, providing a safety margin from the limit of 29 pL/cell/ day, was used in a previous work for the cell growth in the seed bioreactor at pilot scale (Schwarz, Gomis-Fons, et al., 2022). Our group has systematically worked with ActiPro supplemented with the feed concentrates Cell Boosts 7a and 7b of Cytiva, which (together) contain glucose, amino acids, and other proprietary components for the production process (Schwarz, Gomis-Fons, et al., 2022;Schwarz, Mäkinen, et al., 2022;Zhang, Schwarz, et al., 2020).
In view of the main purpose to optimize the amino acid concentration in perfusion medium, a first screening experiment, Screening-exp#1, with nine different media varying only in the amino acid composition, was performed in semi-perfusion CHO cultures.
The amino acid concentrations of these nine media, p1-p9, were different cocktails obtained from previous in-house testing, chosen to provide different metabolic behaviors as shown in Supporting Information: Table S1. Media p1-p9 were mixed by supplementing N0 medium. CHO cells, inoculated from the same preculture, were cultured for 10 days in semi-perfusion mode to mimic steady-state perfusion cultures. The seeding cell density was 18 × 10 6 cells/mL and the medium was daily renewed, which corresponded to a CSPR of 55 pL/cell/day. Spent medium analysis was performed to identify differences in the metabolic profiles of the CHO cell line in semiperfusion mode triggered by changes in the amino acid composition.
For each condition, that is, media p1-p9, the last days of semiperfusion culture were taken into account to calculate the average cell-specific consumption rates of the amino acids. The aim of the present study was to focus on a few relevant nutrients rather than have a systematic overall optimization, which would have been overwhelming. Asparagine (Asn), serine (Ser), and leucine (Leu) had the largest variations in cell-specific consumption rates, and to a lower extend aspartate (Asp) as well, while only minor variations, typically not exceeding 0.1 pmol/cell/day, were observed for the other amino acids, see Supporting Information: Figure S2. This led to the selection of the amino acids Leu, Asn, and Ser for further optimization of the amino acid composition.

| Design for optimization of selected components
For the investigation of the effect of Asn, Ser, and Leu, the concentrations of the other amino acids in the perfusion medium had to be adequately chosen. For this, it was decided to adopt concentrations ensuring that limitation would not occur. This was determined with a targeted feeding ("TAFE") approach (Särnlund et al., 2021;Zhang, Schwarz, et al. 2020). In this approach, a target cell-specific consumption rate of a given nutrient is decided and the concentration of this component, for example, amino acid, is calculated according to where AA medium is the concentration of an amino acid to be set in the perfusion medium and q aa target is the target cell-specific uptake rate of the amino acid.
In short, Equation 6 is derived from the mass balance equation of an amino acid in steady-state culture with the assumption that the cell-specific consumption rate of this amino acid is constant and is completely consumed, leading to a residual concentration noll (Särnlund et al., 2021). The application of Equation 6 is counter intuitive and against the way a scientist developing a cell culture process would normally select the concentration of a substrate in the culture medium. As a matter of fact, the general practice is to select a concentration to feed a substrate based on experience, knowledge of the cells, or trial/error, to monitor the residual concentration of this substrate in culture and then to calculate the cell-specific consumption rate of this substrate in the experimental setting. Using TAFE, a target of the cell-specific consumption rate of the substrate is selected beforehand.
For most amino acids, the q aa target was decided from the maximum cell-specific uptake rate in semi-perfusion cultures at 55 pL/cell/ day measured in Screening-exp#1, see Table 1. From our previous observations, it was known that the amino acid cell-specific consumption rates were typically lower in perfusion cultures compared to semi-perfusion owing to slower metabolism in particular at low CSPR. The selection of amino acid concentrations, AA medium , using q aa target based on measurements of semi-perfusion cultures at CSPR 55 pL/cell/day would therefore avoid the risks of limitations of these nutrients at CSPR 17 pL/cell/day in perfusion mode. The concentrations of Asn (5.1 mM), Ser (5.1 mM), and Leu (4.0 mM) were chosen based on much lower q aa target to study their effect by individual supplementation in the perfusion media in the DoE experiment of medium optimization, DoE-exp#2. Notice as well that the prototype medium, used for the study, was based on ActiPro with reduced concentrations of amino acids, which gave constraints for the selection of the lower limit of the amino acid concentrations.
In this experiment, the amino acid concentrations listed in Table 2 were present in medium M0, which was mixed from N0 and was also supplemented with 55 mM glucose. A concentration of 55 mM glucose corresponded to a targeted cell-specific glucose consumption rate of 1 pmol/cell/day for a CSPR of 17 pL/cell/day according to Equation 6, adapted for glucose, in line with our previous study Zhang, Schwarz et al. (2020). For the experiment DoE-exp#2, the target VCD was 55 × 10 6 cells/mL at a perfusion rate of 1 vvd.
Besides the effect of amino acids, the carbohydrate source can have an influence on the antibody production as previously studied (Zhang et al., 2019;Zhang, Schwarz, et al., 2020;. For this reason, the effect of a fourth substrate was studied, galactose (Gal), supplemented in addition to glucose.
According to our previous investigations, this latter carbohydrate can be efficiently taken up by the cells in presence of glucose despite the much higher uptake rate of this latter compared to Gal, thanks to A simplex design was adopted for this experiment since this approach, applied by others, is known to be well-suited for mixture design (Didier et al., 2007;Jordan et al., 2013;Kim & Lee, 2009;Liu et al., 2015). The design was a simplex-centroid including the four media (M1-M4), which resulted in 15 mixtures (N1-N15) with three center points (N15-N17), and is presented in Table 3.

| Medium optimization in perfusion MBRs-experiment DoE-exp#2
For experiment DoE-exp#2, the mixtures were tested in nine steady- During 17 days of perfusion, no significant retention of mAb by the PES filter was observed. The volumetric productivity Q p , presented in Figure 3f, was computed as factor oḟ * H P HT 2 , excluding product loss by cell bleed. It was observed that the culture performance in terms of growth and productivity of byproducts and mAb varied significantly between the different media mixtures (Figure 3 and Table 4).
T A B L E 1 Determination of the amino acid concentrations in the perfusion medium by TAFE approach.
Note: "Measured uptake rate" range of amino acid cell-specific consumption rates in CHO semi-perfusion cultures at 55 pL/cell/day from Screening-exp#1 (production is listed as 0 or not given); "Target rate" selected target specific consumption rates of the amino acids, q aa target , mostly taken as the highest observed rates for the perfusion medium; "Concentration" concentrations of amino acids in the perfusion medium calculated using Equation 6.
Abbreviations: CHO, Chinese hamster ovary; GS, glutamine synthetase; TAFE, targeted feeding. a Essential in GS-CHO cells under glutamine omission.
b Produced amino acids.
T A B L E 2 Concentrations of glucose, asparagine, serine, leucine, and galactose in media M0-M4 used for the DoE-exp#2 experiment to study the effect of these components on the perfusion culture performances. As a result, the N-glycosylation profile of mAb was also varied in the different media mixtures (  (Figure 4a), DO (Figure 4b), and the perfusion rate (Figure 4c) are shown for two representative cultures. The data showed stable profiles for these process parameters, which were comparable to what can be achieved in traditional STBRs. The OD sensor, which was used in a control loop to stabilize the VCD, served a similar purpose to conventional biomass sensors applied in perfusion (Figure 4d).
Deviations in the perfusion rate and OD typically occurred every 24 h due to the sampling. Just after the sampling, bolus injections of the respective perfusion media, N1-N15, (of a volume equal to the sample volume) were introduced into the MBR such that the working volume remained unchanged after sample removal. The additional media injections to equalize the working volume were neglected for the perfusion rate calculation, resulting in a small systematic error in the perfusion rate of 5%. Maintaining a perfusion rate within ±5% is quite satisfying given that the working volume is very small. Notice that due to the 5% error in the perfusion rate, the calculations of the productivities have a small systematic bias of 5%, that is, underestimation, which is the same for all the conditions, so a comparison of the different conditions with each other is not affected. Notice that the extrapolation of the MBR results to a stirred bioreactor is a little bit biased by a value of 5%, which is very small.

| Selection of optimized medium
Data analysis by regression modelling was employed for the medium design optimization using the data of experiment DoE-exp#2. Partial least square (PLS) regression was applied to fit the models, where experiment DoE-exp#2 data in Table 4 Figure S3 for cases of constant fraction of M1 or M3 set to 0 or 0.25. No valid models with sufficient predictive power were obtained for the mAb titer and its productivity, and therefore these performance indicators were not considered for decision-making directly based on the DoE. Of course, the medium optimization aimed at maximizing the mAb volumetric productivity Q p . An important aspect is that the growth rate together with the mAb titer affect directly Q p . As previously mentioned, during the production at steady state, the cell bleed rate is equal to the growth rate, and the mAb present in the cell bleeds is lost. The volumetric productivity, which takes into account this loss, is thus lower for high cell growth for a given mAb concentration in the bioreactor. A low growth rate is thus desirable in steady-state perfusion to prevent this significant loss of mAb from cell bleeding. In this study, an objective of growth rate between 0.2 and 0.25 day −1 was selected as compromise of stable Note: The numbers in columns M1-M4 give the relative proportion of the four media in the medium mixtures N1-N17 obtained with a simplexcentroid mixture design where for each medium mixtures (N1-N17), the media M1-M4 are supplied in fractions between 0 and 1 such that all the fractions of the medium mixtures (N1-N17) always add up to 1 for a given condition; the cultures were carried out in the MBR and numbered C1-C9, listed in column "Run" with colors alike the line colors used in Figure 3; all the runs except C9 included sequentially two steady states corresponding to two different media, performed at Day 3-Day 10 (#1 in column "Steady state") and at Day 11-Day 17 (#2 in column "Steady state").

SCHWARZ ET AL.
| 2531 cell growth to ensure a high viability and low product loss from cell bleeding. Q p objective was set to a minimum of 575 mg/L/day for steady-state perfusion at 55-60 × 10 6 cells/mL, representing the upper quartile (75th percentile) of the data set. The objective for the viability was set to >95%, a limit achieved in all the media N1-N17 (Table 4).
It can be observed from Supporting Information: Figure S3 that the case of M1 fraction set to 0 (third column of plots) resulted in the lowest predicted lactate concentration in culture (between 2 and 14 mM) in comparison to the other cases. Furthermore, M1 fraction set to 0 led also to a wider feasible area of G0-level glycans >36%, which was taken as objective for this parameter. The N-glycosylation profile of the reference originator mAb had a different pattern compared to any profile observed in the tested conditions, with a high percentage of G0-level glycans (G0 and G0F) of around 49% (Joshi & Rathore, 2020). We selected the median for the lower limit of the relative amount of G0-level glycans (>36% G0 and G0F). It should be noted that the aim of this study was not to fully recreate the N-glycosylation profile of the originator mAb, therefore, the acceptable range for the N-glycosylation was arbitrarily set. It was thus decided to set M1 fraction to 0, that is, to exclude M1. This  Table 4.
A sweet spot plot was generated to support the medium selection using Modde software based on the above-mentioned objectives for the growth rate, the G0-level, and the ammonia concentration in the bioreactor, see recapitulated list of objectives in Table 5. Figure 5 shows the medium mixtures obeying these objectives where the legend of Figure 5 details the different areas.
Constraints on the growth rate narrowed the feasible range to a band with a maximum proportion of ≈ 0.5 fraction of M2 and of M3 and ≈ 0.7 fraction of M4 (between the contours "Growth Max" and "Growth Min" in Figure 5). To further constrain that region, the constraint of G0 level >36% was visualized as the region to the left of the contour "G0 Min" and the constraint of residual ammonium concentration <1.5 mM as the area to the right of the contour "Ammonia Max." The resulting sweet spot of the constraints for the growth rate, the G0-level and the residual ammonia concentration in the bioreactor is shown as a green area in Figure 5. Mixture N10 with an equal ratio of M3 to M4 lied within this feasible area and was therefore selected as the optimum medium for the subsequent validation runs. Figure 5b represents    cell growth leading to loss of mAb in the cell bleed and the mAb titer.
It can also be observed in this figure some tendency of the mAb volumetric productivity Q p to decrease with the growth rate, while the mAb titer increases for values above 0.35 day −1 .

| Validation run of optimized perfusion medium-Validation-exp#3
The validation run with the selected medium N10 was performed in several bioreactor settings. Steady-state perfusion experiments were performed, where constant VCD and perfusion rate of 55-60 × 10 6 cells/mL and 1 vvd, respectively, were maintained for a period of 10 days in a STBR with an ATF cell retention system, in a STBR with TFF and in duplicate MBR runs. Semi-perfusion runs with duplicate spin tube cultures were performed in parallel to the STBR and MBR experiments, aiming for the same average VCD, with one medium exchange every 24 h. The results of the culture performances are shown in Figure 6. From Day 2 of the cultures, the VCD stabilized in the targeted range and the viability was maintained above 95% (Figure 6a). Considering the last 5 days of each culture, the growth rates were fairly similar in the STBR (0.18 ± 0.04 day −1 ) and MBR runs (0.21 ± 0.03 day −1 ) ( Figure 6b). The average growth rate in semi-perfusion cultures was considerably higher with 0.33 ± 0.03 day −1 . Surprisingly, lactate concentrations varied to a greater extent even between the ATF and TFF cultures with less stable concentrations in the latter (Figure 6c). Ammonium was consumed with 1.17 ± 0.19 mM and 0.82 ± 0.05 mM residual concentration in the STBR and MBR processes, respectively, while in the semi-perfusion system it switched from consumption to production on day 5 (Figure 6d). It is probable that in spin tube BR, where the medium change is comparable to batch refeed mode, the glucose consumption rates are initially higher from the increased availability of glucose after resuspending the cells in fresh medium, causing a periodic depletion of glucose and switch to lactate consumption.
After the first few days of culture adaptation, the residual lactate level in these semi-perfusion cultures reached low levels near 0 mM. This was accompanied by an increase in amino acid consumption, which could be a compensation for the reduced availability of sugars or lactate, so that the cells had to generate energy from other sources, such as the amino acids, which then resulted in secretion of ammonium. A similar observation was made previously when the same cell line was cultured in perfusion mode with very low sugar feeding .
Spent medium analysis in semi-perfusion cultures revealed a higher depletion of amino acids compared to continuous perfusion cultures (Figure 6g), which may be explained by the different kinetics in batch and perfusion mode. These observations might also explain the higher cell growth and cell-specific productivity in semi-perfusion cultures. The difference in the mAb titer between the TFF and ATF culture was due to higher product retention by the TFF system (Figure 6e), however, the cell-specific productivity was not affected (Figure 6f). With respect to the product quality T A B L E 4 (Continued) VCD (10 6 cells/mL) 99.0 ± 0.54

Note:
The N-glycosylation of mAb was analyzed for the last day of each condition.
Abbreviations: DoE, design of experiment; VCD, viable cell density.
attributes, only minor differences were observed for the Nglycosylation pattern of mAb, with slightly higher similarity between STBR and MBR compared to the semi-perfusion system and STBR (Figure 7a). Moreover, the charged variant distribution of mAb was influenced by the cell culture system. Both, semiperfusion and MBR system did not accurately mimic the charged variant distribution of mAb produced in the STBR (Figure 7b). This is possibly because the residence time of mAb was slightly higher in the STBR due to partial filter fouling of the hollow fibers. We have been previously observed an increase in acidic variants of the present mAb from higher residence time (data not shown).
From a comparison of the results obtained in both scale-down models, the MBR demonstrated a better performance than the semiperfusion system in predicting the cell culture behavior of an STBR perfusion process. Furthermore, the expected responses from the medium optimization study were validated with relatively high accuracy in the STBR and MBR cultures.
To further demonstrate that our medium development strategy is applicable to very high cell density perfusion cultures exceeding F I G U R E 4 Online data of two representative runs: C1 (light gray) and C4 (dark gray) in the microbioreactors (see Table 3 for details about C1 and C4 runs); (a) dissolved oxygen, (b) pH, (c) perfusion rate, (d) optical density. Time between the dashed lines indicates the two "3-day steady states" (see Figure 3). The peaks on Day 4 in (b) are due to manual pH corrections after performing pH offline measurements. The peaks in (c) are due to short bolus injections of perfusion medium at high feed rate during sampling events.
T A B L E 5 Feasible medium mixtures (sweet spot generated using Modde software to support the medium selection); target range, model prediction, and results from validation in MBR and STBR.  to support a very high cell density. As a result, perfusion process development aims at limiting the medium demand by identifying the minimum CSPR (Konstantinov et al., 2006). Considering the report about a common framework for integrated and continuous F I G U R E 5 (a) Sweet spot plot identified from the model using experiment design of experiment-exp#2 data showing the feasible mixtures under variation of the relative fractions of M2, M3, and M4 while M1 = 0, with primary constraints on G0-level glycans (where the area to the left of the contour "G0 min" is such that G0 level ˃36%) and growth rate (where the area between the contour "Growth Max" and "Growth Min" restrains the zone of growth rate between 0.2 and 0.25 day −1 ), and a secondary constraint on the maximum ammonium concentration (where the area to the right of the contour "Ammonium Max" is such that the ammonium concentration in the bioreactor is <1.5 mM). The turquoise area indicates the feasible mixtures obeying the objectives of the growth rate and the G0-level and includes a smaller green area, which corresponds to the zone where the objective for the ammonia concentration is respected. The green area is thus the target zone obeying all the objectives of G0-level glycans, growth rate, and ammonia concentration in the bioreactor. (b) Monoclonal antibody (MAb) titer and productivity as a function of the cell growth rate for the different media N1-N17, showing that the mAb productivity of medium N10 is the highest, despite not corresponding to the highest mAb titer.
F I G U R E 6 Comparison of perfusion mode (left panels) with MBR cultures performed with biological duplicate (n = 2) (black circles), ATF-STBR culture (orange diamonds), TFF-STBR culture (blue squares), and semi-perfusion cultures (right panels) in spin tube BRs performed with biological duplicate (n = 2) (black triangles) using the optimized perfusion medium in experiment Validation-exp#3; (a) viable cell density (straight lines), average cell density in spin tube cultures in 24 h interval (dotted line) and viability (dashed lines), (b) growth rate, (c) lactate concentration, (d) ammonium concentration, (e) mAb harvest titer, (f) mAb cell-specific productivity-error bars represent the standard deviation for n = 2 in panels a to f; (-g) amino acid concentrations of N10 medium and average amino acid concentrations over 5 days in ATF-STBR culture, TFF-STBR culture, MBR cultures (biological duplicate n = 2), and semi-perfusion cultures in spin tube BRs (biological duplicate n = 2)-error bars represent the standard deviation. ATF, alternating tangential flow filtration; BR, bioreactor; mAb, monoclonal antibody; MBR, microbioreactor; STBR, stirred-tank bioreactor; TFF, tangential flow filtration.
biomanufacturing (Coffman et al., 2021), we set a limit on the CSPR of 17 pL/cell/day for the optimization of the perfusion culture.
The present study was initiated by studying nutritional requirements of the CHO cell line by a screening in semi-perfusion cultures (Screening-exp#1). With this approach, the cell requirements for important nutrients, the amino acids, were identified and it was observed that Asn, Ser, and Leu were most often the most consumed using different amino acid cocktails, and as a result these were selected for further optimization. The VCD was kept relatively low at 20 × 10 6 cells/mL to alleviate the operations at this early stage. It allowed as well working with suboptimal medium composition since a high CSPR of 55 pL/cell/day was used. The observed amino acid cellspecific uptake rates at this CSPR were translated into concentrations to support a perfusion culture at our target CSPR of 17 pL/cell/day, that is, through the delivery with amounts not resulting in nutrient limitations. TAFE was demonstrated for sugar delivery  and this approach was demonstrated as well in a transfer from fed-batch to perfusion culture by adjusting the concentration of amino acids in perfusion medium to match the F I G U R E 8 Very high cell density perfusion runs in the microbioreactor (n = 2 as one culture run for 11 days and the other one for 14 days) with optimized perfusion medium at a perfusion rate of 2 vvd in experiment Validation-exp#4; (a) viable cell density (straight line) and viability (dashed line), (b) Monoclonal antibody titer in bioreactor (full circles) and harvest (empty circles), (c) volumetric productivity. Perfusion was initiated with 1 vvd and after 1 day increased to 2 vvd using standard ActiPro. From Day 1 to 3, the ratio of optimized medium to standard ActiPro medium was changed from 0% to 50% and then 50% to 100%. From Day 4, the cell density was maintained between 110 and 120 × 10 6 cells/mL. Error bars represent the standard deviation as value of n = 2.
cell-specific consumption rates with those observed in fed-batch (Särnlund et al., 2021). The same concept was here adopted for the transfer from high-CSPR semi-perfusion culture to low-CSPR in perfusion bioreactor. Notice that the selected low levels of concentrations for Ser and Leu, 5.1 mM and 4.0 mM, were high in comparison to the residual concentrations of 0.5-1 mM listed as growth inhibitor for fed-batch cultures in Pereira et al. (2018). However, these corresponded to low cell-specific consumption rates at high cell density perfusion with low perfusion rate and had no influence on the viability and the cell growth, as can be seen in Figure 3 and shown that time-consuming ramp-up phases could be reduced by at least 5 days through high cell density inoculation between 30 and 60 × 10 6 cells/mL, still requiring a relatively low total cell number in the inoculum due to the very small working volume of 2 mL. This was also appealing because, in the present study, the interest was the optimization of the steady-state phase, which is the most important contributor to the production of mAb, rather than optimizing the growth phase. Furthermore, the availability of a true perfusion MBR for optimization work, in contrast to semi-perfusion systems, also increases the likelihood for a successful process transfer to STBR. A potential drawback inherent to the small working volume of the MBR is the small sample volume, which can reduce the application of analytical methods. high Leu concentrations in the perfusion medium, and therefore it was a potential source for growth inhibition in the present study.

It was observed in
The overall objective with the DoE was to identify a medium formulation with improved performance for steady-state perfusion.
The observed responses with respect to the tested mixtures was taken into consideration for the medium design. Regression models were estimated for each individual response as a function of the proportion of the four media to support medium optimization.
Subsequent to the optimization, validation runs were performed in STBRs and MBRs, experiment Validation-exp#3, demonstrating an overall similar performance on both scales that was also in agreement with the predicted responses from the DoE study. These results were further confirmed at cell densities of 110-120 × 10 6 cells/mL, illustrating that the present medium development strategy was applicable to densities exceeding 10 8 cells/mL in Validation-exp#4 experiment. The diverging cell behaviors observed in the semiperfusion runs in comparison with perfusion culture highlights the importance of using a small-scale perfusion system for fine-tuning of the medium formulation in process development. For example, the growth rate and metabolite concentrations were not accurately captured by the batch refeed culture system. Furthermore, such a system is limited in the maximal achievable cell density, and to our

| CONCLUSION
A DoE-assisted medium optimization strategy was developed to design the concentrations of key nutrients in perfusion medium, where the concentrations of the amino acids were determined based on targeted cell-specific uptake rate approach TAFE. The experimental effort was greatly reduced by employing the novel perfusion MBR Erbi Breez, capable for steady-state perfusion at cell densities of 60 × 10 6 cells/mL for 17 days and possibly longer, as well as for very high cell density perfusion above 10 8 cells/mL for 2 weeks. With the proposed strategy, an optimized medium formulation was obtained for perfusion at a CSPR of 17 pL/cell/day, yielding in reduced growth during steady-state cultivation, reduced secretion of byproducts lactate and ammonium and a mAb N-glycosylation profile in accordance with a specified range.
The present approach, exemplified here for amino acids and sugars, can be applied for component optimization in perfusion media. We believe that perfusion MBRs in combination with the present medium optimization strategy can contribute to very rapid developments of intensified perfusion processes.

ACKNOWLEDGMENTS
The authors would like to thank Erbi Biosystems for providing the microbioreactor system and his support. The authors also show appreciation to Andreas Andersson of Cytiva and to Cytiva for providing the cell culture medium and the cell line. This work was financed by the Sweden's Innovation Agency VINNOVA (diaries nr. 2016-05181 and 2022-03170), Competence Centre AdBIOPRO for Advanced BioProduction by Continuous Processing.