Impact of freeze–thaw processes on monoclonal antibody platform process development

Freezing of cell culture supernatant (CCS) is a standard procedure in process development of monoclonal antibody (mAb) platform processes as up‐ and downstream development are usually separated. In the manufacturing process of mAb, however, freezing is avoided, which poses the question of comparability and transferability from process development to manufacturing. In this case study, mAb CCS from Chinese hamster ovary (CHO) cells is frozen and thawed in a novel active freezing device and subsequently captured by protein A chromatography. Critical quality attributes such as host cell protein (HCP) concentration and soluble mAb dimer shares have been monitored throughout the case study. Furthermore, cryo‐concentration of individual proteins was investigated. The main factors that drive cryo‐concentration are diffusion and natural convection. Natural convection in freezing processes was found to increase at warmer freezing temperatures and thus slower freezing, leading to higher concentration gradients from top to bottom of a freezing chamber. The freeze concentration was dependent on protein size and correlated to diffusivity, where smaller proteins are exposed to higher cryo‐concentration. Our results suggest that as a result of freezing processes, large particles based on mAb and specific host cell proteins (HCPs) expressing a certain affinity to mAbs are formed that have to be removed before purification. This leads to a significant improvement in HCP reduction by the protein A step, when compared with reference samples, where twice as much HCP remained in the eluate. Furthermore, HCP and mAb dimer concentrations in protein A eluate were dependent on the freezing temperature. As a conclusion, CCS should be frozen as rapidly as possible during process development to minimize issues of transferability from process development to manufacturing.


| INTRODUCTION
In the biopharmaceutical market, monoclonal antibodies (mAb) are the most important class of proteins to date. Hence, many studies with industrial interest in the improvement of the manufacturing process have been published leading to the establishment of a platform process for mAb purification (Baumann & Hubbuch, 2017).
A typical manufacturing process involves three major steps: cell culture, purification, and formulation. During large-scale manufacturing, each of these steps is performed subsequently, whereas all steps are evaluated individually during process development. Additionally, the up-and downstream manufacturing parts are usually located on one manufacturing site, while filling of the drug product is often done at different locations. Therefore, the product is often frozen to reduce the risk of product loss by microbial growth, foam prevention, and mechanical stress during transportation and hold times in process development (Authelin et al., 2020;Kolhe & Badkar, 2011). However, freeze-thaw (FT) process steps come with disadvantages that might lead to protein activity loss (Bhatnagar et al., 2007) such as cryo-concentration, protein-ice surface interaction (Chang et al., 1996), and cold denaturation (Privalov, 1990).
These FT stresses might lead to protein aggregation (Mahler et al., 2009) and even native aggregate particle formation, which was previously reported for mAb (Telikepalli et al., 2014). Because of the high importance to the drug industry, several studies on FT processes of mAb formulations have been performed (Hauptmann et al., 2019;Miller et al., 2013;Rayfield et al., 2017). Such freezing processes are often categorized by scale and mode of cooling. Actively cooled freezing processes involve a cooling fluid, which is in contact with the container wall, whereas containers frozen in larger freezers are referred to as passive freezing processes. Across all freezing processes, cryo-or freeze concentration occurs due to exclusion of the solutes from ice crystals. This phenomenon is also well described for different solidification processes in various areas such as alloys (Shevchenko et al., 2015). From a macroscopic view, the area of solidification at the freezing front, where crystals grow into solution, is described as a "mushy zone." As a result of freeze concentration, buoyancy-driven flows also known as "natural convection" occur in the mushy zone and the remaining liquid in large scaleprocesses, due to density gradients. In addition to natural convection, diffusion of solutes also leads to cryo-concentration (Butler, 2002), which can be described by Fick's law.
As pointed out initially, up-and downstream process development is often done with hold times in between and process intermediates have to be frozen to increase their shelf life. This step is avoided during the manufacturing process to minimize the risk of product degradation. Despite this major difference in process development and manufacturing, no studies have been presented yet on the transferability of data characterizing unit operations across scales, where different sample preparations (frozen vs. reference) were applied. Therefore, this study investigates the impact of an additional FT cycle before a typical protein A step using cell culture supernatant (CCS) from Chinese hamster ovary (CHO) cells with mAb. Furthermore, characterization of complex freezing processes with multiple proteins is performed to provide a better understanding of freezing processes.

| mAb preparation
CCS of a mAb harvest with a titer of 2 g/L from CHO cells was kindly provided by Byondis. Due to the lack of stability of CCS, handling and storage of CCS is not possible without freezing. Thus, the product was frozen at −80°C post cell removal at the production site, and stored in 1 L bottles until further use. To adjust the sample volume, CCS from 1 L bottles has been thawed in a water bath at 25°C for 2-3 h, aliquoted into 50 ml centrifugation tubes by Corning Life Sciences at 45 ml and frozen at −80°C until further use. Before an experiment, the required number of aliquots have been thawed in a water bath at 25°C, pooled, and filtered with 0.2 µm filters. In total, the harvest was freeze-thawed twice and filtered once before conducting the case study. This may influence the outcome of the study, but mirrors typical handling of process development samples in industry. This said, the significant results obtained in this study highlight the mechanisms occurring in any FT process during sample handling. The twice freeze-thawed harvest before our study will be referred to as the "reference sample."

| FT process
Controlled freezing and thawing was done in a small-scale freezing device designed and manufactured in cooperation with Industrietechnik Salzburg Bilfinger. The freezing container is designed as a hollow tube cooled from the in-and outside. The used scale-down model is designed as a thin slice of a larger scale, that is separated into six wedges by an insulating inlay of polytetrafluoroethylene. A schematic drawing of the freezing device and a sample chamber is displayed in Figure 1.
To reduce boundary conditions, an additional cooling circuit heats the wedge from the bottom at a constant temperature of 0°C.
The detailed process and thermal process behavior are described thoroughly in a previous study on the characterization of freezing processes .
All freezing experiments were performed in triplicates using three separated freezing chambers to account for process variation and assure reproducibility. Seventy-five milliliters of CCS prepared as described above was pipetted into each chamber.
Then, the device was tempered at 5°C for at least 1 h for temperature equilibration throughout the system and the sample bulk. After temperature equilibration, freezing was initiated by lowering the cooling fluid temperature at maximum cooling rate to −60°C to −20°C. The final temperature was held for at least 5 h or overnight.
A core drill with 8 mm inner diameter from Buerkle was used for sampling from the frozen bulk. A 3D-printed lid was put on top of a chamber to assure reproducible sampling at three levels of 8 mm at nine locations with equal radial distances from each other as shown in Figure 1b. In preliminary experiments, radial freeze concentration was found negligible. Thus, samples were taken from two neighboring rows providing an increased number of overlapping samples volumes for improved data resolution. Afterwards, the frozen bulk was thawed by increasing the cooling fluid temperature to 25°C for 1.5 h before homogenization and final liquid sampling using a 5 ml pipette.
If a subsequent protein A capture step was performed, no frozen samples were taken. In this case, the device temperature was lowered to 5°C after thawing to reduce protein degradation while each replicate from a separate chamber was processed resulting in three separate protein A runs per freezing temperature. Therefore, samples were kept inside the cooled freezing device for up to 7 h at 5°C.

| mAb capturing
The mAb was captured from CCS using protein A affinity chromatography. MabSelect Sure was packed and operated with an ÄKTApurifier system (Cytiva). UV extinction at 280 nm, pH, and conductivity were measured throughout the purification. An Omnifit column with 10 mm inner diameter from Omnifit Ltd. was used as column housing. The column was repacked throughout the experiments with column heights ranging from 188 to 203 mm resulting in a column volume (CV) of 14.7-15.9 ml. Sixty-five milliter samples with a titer of approximately 2-3 g/L mAb were loaded onto the column. As the manufacturer states a dynamic binding capacity of 35 g/ml resin, the column was operated well below its maximum capacity. The chromatography was conducted at a constant flow rate of 300 cm/h with PBS, pH 7.4 as equilibration and post-loading wash buffer. Twenty-five millimolar NaAc, pH 5 has been applied as a second wash followed by an elution with 25 mM acetic acid, pH 3 until 2 CV after the end of fraction collection. Product was collected in 15 ml fractions starting from an extinction of 0.2 AU until stopped below 0.1 AU. Fraction collection criteria were chosen based on recommendations by the harvest supplier. Detailed chromatography conditions are listed in the Supporting Information Material.

| Filtration analysis
In general, as some systems showed a high turbidity, 0.2 µm filtration was performed in between every FT and chromatography step to avoid clogging of the protein A capturing columns. A filtration cascade using syringe filters with 1.2, 0.45, and 0.2 µm cut-off from Sartorius was performed to investigate the size range of the particles present. The flow-through was analyzed for particles as described below.
Furthermore, the filter cake retained after a filtration step was investigated to measure the loss of mAb and host cell protein (HCP) through filtration post freezing and thawing. Therefore, Vivaspin 2 filters with a 0.2 µm cut-off polyethersulfone membrane from Sartorius were loaded with a sample volume V load of 2-4 ml and centrifuged at 1200g. In preliminary experiments, 20 ml could be loaded onto the filter until filter clogging occurred. To reduce the influence of membrane fouling with increasing load volume, maximum load volumes of 4 ml were applied. After filtration, the filter membrane was detached from the housing, transferred into 500 µl centrifugation tubes and incubated with 500 µl V dissolve size-exclusion chromatography-high-performance liquid chromatography (SEC-HPLC) running buffer at 700 rpm and 5°C in a thermomixer comfort from Eppendorf overnight to dissolve any retained aggregates. As the filter membrane contained solution in the membrane pores after centrifugation, the measured protein concentration c dissolve in the dissolution buffer was composed of the dissolved filtered particles and the protein concentration from the filtered solution remaining in the filter. Thus, a mass balance over different load volumes V load was used to calculate the aggregate concentration c aggregate in the F I G U R E 1 Freeze-thaw scale down model. (a) An exploded view of the device with two cooling walls in blue. The inlay is used to reduce boundary freezing from the bottom and for volume reduction. The bottom of the device is heated to further minimize freezing from the bottom. (b) A single chamber with sample layers and drill holes in the bulk volume as indicated by the dashed cross-section in (a). Images were adapted from Weber and Hubbuch (2021) analyzed sample. The subtraction of mass balances with different load volumes eradicates the influence of the remaining solution in the pores and leads to the following equation: where m is the slope of a linear regression of c dissolved over V load .
Protein concentrations were determined by capillary electrophoresis, HPLC, and SEC as described below. As proteins tend to adsorb to the filter membrane, the flow-through was re-filtered in a separate experiment for comparison.

| Analytics
As this study aims to mimic typical process development conditions, analytics have been carefully chosen with respect to their application in industry. The most commonly used analytics involve enzymelinked immunosorbent assay (ELISA) to quantify HCP (Wang et al., 2009), SEC for quantification of mAb monomers and aggregates and sodium dodecyl sulfate-polyacrylamide gel electrophoresis for protein size detection and quantification for low concentrated HCPs. While the analytics partially provide redundant data, comparability of redundant results is not always given due to lower limits of detection of the used methods. Additionally, the optical density of solutions can be correlated with the particle number of non-filtered samples. Statistical significance was analyzed using a paired-sample t test.

| Large particles analysis
The optical density was used as an indicator for large aggregates in non-filtered samples. Therefore, 400 µl of the sample was pipetted into a cuvette with 1 cm path length, and the extinction at 600 nm wavelength was measured in a photometer Infinite 200 by Tecan.
2.10 | mAb monomer and aggregate analysis

| Protein concentration
The mAb concentration in post-capture samples was measured with Nanodrop 2000c by Thermo Fisher Scientific. Assuming a negligible HCP content, UV absorption was measured at 280 nm and protein concentration was calculated with an extinction coefficient of

| mAb concentration from CCS using SEC-HPLC
The quantification of mAb monomer and dimer was done using SEC-HPLC in a manner similar to Paul et al. (2014). Spiking of the CCS with concentrated mAb was performed to validate the determination of mAb concentrations from absorption areas of SEC-HPLC chromatograms.
As shown in Figure

| Freeze concentration profiles
The macroscopic freeze concentration in frozen bulks was in-

| Filtration analysis of thawed mAb CCS
After freezing and thawing of CCS, sometimes high turbidities can usually be observed visually. When thawed CCS was filtered using a filter cascade of decreasing pore size, the turbidity gradually decreased as shown in Table 1. The majority of the filtered particles (66%) was bigger in size than 1.2 µm, whereas 24.8% had a size | 3919 between 0.2 and 0.45 µm. Hence, 0.2 µm filtration before chromatography steps is necessary to avoid clogging of the columns. It has to be noted that several pre-syringe filters had to be used after freezing due to the large number of particles blocking the filter pores. This poses the question as to whether product is lost during filtration and critical quality attributes are changed. To analyze the particles filtered from a solution, a redissolution step of the retentate was performed. After a centrifugal filtration step, the membrane remains wet with soluble proteins in the retained solution. This leads to incorrect concentration measurements when incubating the filter membranes in redissolution buffer. Therefore, it is necessary to filter different filtration volumes, which is shown in Figure 5a, with exemplary raw data of replicates from individual freezing experiments.
Under the assumption of negligible membrane fouling, the redissolved protein concentration c dissolve should be proportional to the volume filtered V load . Thus, a linear regression was performed for each experiment. This regression reduces measurement errors occurring due to standard deviations and membrane variation. Additionally, the concentration contribution of the solution retained in the membrane pores can be calculated from the y axis intercept.
When comparing the redissolved protein concentrations from CCS frozen at −40°C and −20°C, the regression slopes decrease from 11.7 to 1.9 µg·ml −1 ·ml −1 (µg protein per ml V dissolve and ml V load ) with statistical significance (p = 0.04). The particle concentrations calculated from the slopes are displayed in Figure 5b. More particles containing mAb and HCP are filtered on average when the solution is frozen at −20°C. While the concentration of the total protein content and the HCP in the particles could be reduced when lowering the freezing temperature to −40°C, the measured mAb concentration did not change significantly. Only a minor reduction in dimer shares when decreasing the freezing temperature was found on average.
Adsorption of the proteins to the filter membrane did occur, but on average, it was always below the smallest concentration measured.
The measured mAb concentration was approximately one order of magnitude lower than the concentration of HCP and total protein.

| Capturing of freeze-thawed mAb CCS
CCS from CHO was frozen at −40°C, −30°C, and −20°C to investigate the influence of freezing and thawing on critical quality attributes. As particles are formed post FT steps, filtration is necessary to avoid column clogging. When measuring the mAb aggregate and HCP content post FT and filtration, no significant changes were observed. HCP concentrations varied within 101 ± 6% and 117 ± 6% of the initial concentration without a trend regarding freezing temperature. The average initial HCP concentration was 1.45 g/L. The average mAb aggregate content post FT varied from 105 ± 2% to 97 ± 5% of the initial value. Although no significant difference was found, mAb aggregate shares post FT increased on

T A B L E 1 Turbidity reduction with filtration cascade
Filter pore size (µm) Turbidity (mAu)

Turbidity reduction (mAu) (% of total)
Non-filtered 72.7 ± 0.75 n.a.  Therefore, aggregate shares should not be as high as 75% in CCS as reported by Paul et al. (2014).

| Freeze concentration profile
Our finding of macroscopic freeze concentration for elevated freezing temperatures is a well-described phenomenon in slow freezing processes due to elevated temperatures or passive cooling (Hauptmann et al., 2019;Kolhe & Badkar, 2011;Reinsch et al., 2015). processes (Rodrigues et al., 2011) leading to reduced freeze concentration. Due to density gradients at the ice front, natural convection occurs causing high concentrations in the bottom layer . Recent simulations by Geraldes et al. (2020) report velocities in the area of up to 1 mm/s when freezing at −10°C.
Higher density gradients and slower freezing processes, therefore, promote an increased natural convection at high freezing temperatures. The CCS in this study contained different contaminants besides the mAb at 2 g/L such as media components and HCPs. These contaminants are also freeze-concentrated leading to a higher density gradient compared with an mAb solution at a similar concentration, such as final fill formulations. Under the assumption of comparable partition coefficients, natural convection will be more caused by density gradients due to freeze concentration and temperature differences leading to a convective layer in the gravitational direction at the freezing front (Rodrigues et al., 2011;Vynnycky & Kimura, 2007). The convection induces a circular motion, dragging down the solutes to the bottom and along the bottom of the freezing container in front of the freezing boundary, where it settles due to the higher density. Meanwhile, non-concentrated liquid from the center is transported to the freezing boundary, reducing the entrapped solute concentration and thus increasing inhomogeneity in the frozen bulk. On the one hand, natural convection is promoted by high concentration gradients leading to faster velocities at the freezing boundary. On the other hand, the convection at the freezing front is inhibited by crystallization such as dendritic ice formation (Miller et al., 2013). The density anomaly of water adds additional complexity to the mechanism. This complex behavior has been modeled (Geraldes et al., 2020;Ramesh et al., 2021) using the enthalpy-porosity method and the Carman-Kozeny equation to describe velocity through porous media (Kast et al., 2010). Simulations show flow profiles thick as several millimeters at the beginning, which become thinner and slower over time (Geraldes et al., 2020). Smaller molecules with higher diffusivity might be able to diffuse further into the convective layer. Diffusive mass transport might be too slow for molecules to diffuse beyond the boundary layer. Hence, small proteins are exposed to faster drag velocities leading to a higher bottom freeze concentration. Experimental studies of flow profiles during solidification processes show the importance of such flow profiles (Shevchenko et al., 2015), which was also investigated by Geraldes et al. (2020), who simulated the effect of varying mushy zone porosities. Furthermore, the hypothesis is supported by an early study on freeze concentration, where buffer components were freeze-concentrated whereas the lactose dehydrogenase concentration was equal across the frozen bulk (Chen et al., 2001). Other studies have not found significant differences when comparing stabilizing formulation agents such as small buffer molecules and proteins like Roessl et al. (2014), who evaluated freeze concentration in an actively cooled small-scale model. The presumably low height of the container might have reduced natural convection and therefore minimized diffusion-based freeze concentration. The results by Kohle and Badkar (2011) show the mAb aggregate freeze concentration to be similar to that of the monomer, which can be explained by the minor variation in the diffusion coefficient, because of the relative large size of the proteins, which is above 150 kDa.

| FT of CCS leads to temperature-dependent formation of particles containing mAb and HCP
If the CCS was frozen, particles were formed during the FT step, that consist of both mAb and HCP as shown in Figure 6b. On average, all analytics show a reduction in particle formation with decreasing freezing temperature. As lower freezing temperatures lead to faster freezing processes, they induce less freeze stress and therefore reduce the rate of aggregation (Wöll et al., 2019). Furthermore, lower freezing temperatures lead to reduced freeze concentration as discussed before. Such particle formation might be induced by CHO HCPs such as protease cathepsin D (Bee et al., 2015) and thus, should be separated from the product as soon as possible. Furthermore, mAbs also form native aggregates under FT stress (Telikepalli et al., 2014).
The relatively low mAb concentration in the dissolution buffer can be explained by the presence of particles in the dissolution buffer. While ELISA and capillary electrophoresis are able to handle nonsoluble particles by dilution, wash steps or denaturation, HPLC methods require sample filtration and the use of a pre-column filter.
Furthermore, ELISA and capillary electrophoresis are sensitive methods whereas the HPLC operates at the lower limit of the detection and is affected by model limitations as discussed earlier leading to relatively high standard deviations. Finally, capillary electrophoresis indicated a lower total protein concentration after −40°C FT than the measured HCP concentration.
Protein adsorption to filtration membranes is a common issue (Birk et al., 1995) and therefore has to be accounted for at such low concentration levels. The occurring membrane adsorption was lower than the measured particle concentration and thus can be interpreted as the background noise.

| Freezing influences critical quality attributes of mAb capturing
If CCS from CHO cells is frozen before capturing, significant changes in HCP and aggregate shares can be expected in the protein A eluate as depicted in Figure 6. HCP levels in the eluate are generally lower after a FT step followed by a necessary filtration compared with a reference sample. At first glance, the HCP concentration might be generally decreased by the particle removal, leading to an overall reduction of HCP in the protein A load and subsequently the eluate.
However, the HCP content from an aliquot, after the FT step and after filtration did not change significantly. Thus, a hypothesis is suggested, where the changes in the protein A eluate may arise from specific HCPs with high mAb affinity due to protein interactions.
Looking at mAb capturing processes, HCP co-elution with mAb from protein A occurs due to protein interactions between HCPs and the mAb-protein A complex (Bee et al., 2015;Nogal et al., 2012;Zhang et al., 2016). Hence, co-eluting HCPs show increased affinity toward mAbs. If FT stress is exerted on the CCS, the stress favors particle formation of these particular high-affinity HCPs and the mAb because of their affinity. As discussed earlier, warmer freezing temperatures exert stronger freeze stress causing increased particle formation. These particles contain small amounts of mAb and highaffinity HCPs that are removed before the protein A capturing. As a result, freezing and subsequent filtration of CCS decrease the HCP concentration in protein A eluate. Furthermore, the FT step induces mAb aggregation leading to higher aggregate shares (Telikepalli et al., 2014). As a result, slightly stressed and fast-frozen samples at −40°C show higher dimer shares compared with reference protein A eluate. However, the soluble dimer aggregates might form insoluble oligomers that are removed by a filtration step. As higher FT stress is applied at warmer freezing temperatures, larger particles are formed (Hauptmann et al., 2018), which might be caused by higher concentrations (Roefs & De Kruif, 1994