Virus harvesting in perfusion culture: Choosing the right type of hollow fiber membrane

The use of bioreactors coupled to membrane‐based perfusion systems enables very high cell and product concentrations in vaccine and viral vector manufacturing. Many virus particles, however, are not stable and either lose their infectivity or physically degrade resulting in significant product losses if not harvested continuously. Even hollow fiber membranes with a nominal pore size of 0.2 µm can retain much smaller virions within a bioreactor. Here, we report on a systematic study to characterize structural and physicochemical membrane properties with respect to filter fouling and harvesting of yellow fever virus (YFV; ~50 nm). In tangential flow filtration perfusion experiments, we observed that YFV retention was only marginally determined by nominal but by effective pore sizes depending on filter fouling. Evaluation of scanning electron microscope images indicated that filter fouling can be reduced significantly by choosing membranes with (i) a flat inner surface (low boundary layer thickness), (ii) a smooth material structure (reduced deposition), (iii) a high porosity (high transmembrane flux), (iv) a distinct pore size distribution (well‐defined pore selectivity), and (v) an increased fiber wall thickness (larger effective surface area). Lowest filter fouling was observed with polysulfone (PS) membranes. While the use of a small‐pore PS membrane (0.08 µm) allowed to fully retain YFV within the bioreactor, continuous product harvesting was achieved with the large‐pore PS membrane (0.34 µm). Due to the low protein rejection of the latter, this membrane type could also be of interest for other applications, that is, recombinant protein production in perfusion cultures.


| INTRODUCTION
Viral vaccine and viral vector production can be intensified by cultivating animal cells in perfusion mode. The increased cell concentration allows for higher virus titers. To retain cells in the bioreactor vessel, cell retention devices are required that are typically classified by their physical separation principle such as filtration, sedimentation, ultrasonic fixation, or dielectrophoretic exclusion (Castilho & Medronho, 2002). Due to their scalability, simplicity, and efficient cell retention, hollow fiber-based systems are today widely applied for manufacturing of recombinant proteins (Bielser, Wolf, Souquet, Broly, & Morbidelli, 2018). In addition, there is a growing interest for their use in the production of viral vaccines (Gallo-Ramirez, Nikolay, Genzel, & Reichl, 2015;Tapia, Vázquez-Ramírez, Genzel, & Reichl, 2016). With increasing cell concentrations, larger quantities of viruses and other particles are typically released into the medium. In contrast to recombinant protein production, virus infection triggers cell apoptosis, which results in cell degradation and lysis increasing the overall burden of impurities. Besides virus particles and extracellular vesicles, high amounts of DNA and proteins can accumulate. During continuous harvesting this can cause pore narrowing and eventually membrane blockage. In the recent years, numerous studies have reported on such unwanted filter fouling even when large-pore hollow fiber membranes have been applied (Bolton & Apostolidis, 2017;Genzel et al., 2014;Nikolay, Castilho, Reichl, & Genzel, 2018;Nikolay, Léon, Schwamborn, Genzel, & Reichl, 2018;Walther, McLarty, & Johnson, 2018;S. Wang et al., 2017). Product retention was in each case correlated to filter fouling, but a systematic characterization of membrane properties and the performance of retention devices in virus particle harvesting is still missing.
Membrane fouling is studied intensively for downstream processing (DSP), but only a very limited number of studies was performed regarding the use of membranes in upstream processing, that is, for perfusion cultivations. However, product retention has gained more attention as novel production systems aim towards process intensification and continuous biomanufacturing. Three mechanisms are mainly relevant for filter fouling leading to a reduced transmembrane flux and increased membrane resistance (Trzaskus, de Vos, Kemperman, & Nijmeijer, 2015): (1) Internal fouling: adsorption of membrane-compatible particles to the filter material leading to pore narrowing (particle size < pore size); (2) Partial or complete pore blocking: steric pore clogging with particles or agglomerates (particle size~pore size); (3) Gel/cake layer formation: additional solute layer formation of larger particles on top of the membrane by adsorption and subsequent compression by smaller particles (particle size > pore sizes).
While internal fouling typically narrows pore channels, pore blockage, and cake layer formation equally contribute to a reduction in the effective membrane cutoff leading to membrane blockage and filtration termination. The fouling behavior of a hollow fiber module is closely associated with the properties of the membrane, in particular, its pore size distribution, porosity, surface and material roughness, and inner membrane surface charge. Overall properties do not only depend on the specific membrane material used, but also on fabrication procedures and postmodifications (Cornelissen, 1997;Rana & Matsuura, 2010;Ulbricht, Richau, & Kamusewitz, 1998). Due to the high complexity of cell culture processes (e.g., large variation in particles sizes, different surface charges and concentrations, and diverse transport properties), the description of fouling is typically limited to (semi-)empirical models.
Besides the choice of the membrane, specific operational strategies can be established to minimize the risk of filter blockage that take additionally into account the shear sensitivity of animal cells (Futselaar, 1993). First, concentration polarization and boundary layer resistances should be reduced to increase mass transfer coefficients. This can be achieved by increasing the cross-flow velocity (resulting in higher Reynolds numbers) in the filter lumen (e.g., higher flow rate and smaller hollow fiber diameter) or by reducing the transmembrane flux (e.g., lower permeate flow rate and increased membrane area). Another option is the inversion of the tangential flow filtration (TFF) direction resulting in an alternating (bidirectional) tangential flow (ATF). At a given frequency using high flow pulses, this increases the Reynolds number and potentially vortex formation so that foulants are removed more effectively. Second, hydraulic backflushing can be considered by reversing the permeate flow direction across the membrane. This can lift loose deposits on the membrane surface (Hiller, Clark, & Blanch, 1993;Kelly et al., 2014). Likewise, fast inversions of the feed flow direction, as described for ATF systems, can be applied. Thereby, membrane sections along the fiber change periodically the flow direction across the membrane facilitating continuous backflushing with each pump cycle (Radoniqi, Zhang, Bardliving, Shamlou, & Coffman, 2018; Figure 1). Although these hydraulic cleaning methods can be effective, they may increase the shear stress on cells and reduce the net flux. Therefore, it is much more favorable to select a membrane where little fouling occurs, and a stable flux can be easily maintained to reduce the number of hydraulic cleaning steps.
In this study, we investigated different membrane materials and their properties for continuous virus particle harvesting via the permeate for perfusion cultivation. To cover a large variety of different commercial hollow fiber membranes, polyethersulfone (PES), modified PES (mPES), polysulfone (PS), mixed ester (ME, consisting of cellulose acetate and cellulose nitrate), and polyethylene (PE) membranes were tested. If available, two pore sizes (based on nominal cutoff) were investigated to either retain or harvest the virus particles over the cultivation period and to understand filter fouling in dependence of the pore size. This resulted in a sample set of eight hollow fiber modules.
We first characterized different hollow fiber membranes with respect to their potential fouling behavior. In a second step, we tested the membranes in TFF operation for filter fouling and virus particle harvesting. For this, suspension-adapted baby hamster kidney (BHK-21 SUS ) cells were cultured in a bioreactor with an external NIKOLAY ET AL. | 3041 TFF cell retention device (recirculation loop), and the cells were subsequently infected with yellow fever virus (YFV;~50 nm). Filter fouling was monitored in real time using transmembrane pressure sensors, and virus particle, DNA, and protein concentrations were measured in the permeate flow to relate membrane structure measurements to process performance.

| Hollow fiber membranes
Eight commercial hollow fiber membranes (Table 1) were characterized and tested for filter fouling in unidirectional TFF operation.

| Pore size distributions
To determine pore size distributions, a dry single hollow fiber (50 mm length) was potted with a hot glue gun into a PE tubing (5 mm inner diameter [ID]). The end of the fiber was closed with glue and subsequently wetted with the pore-filling liquid fluorinert FC-43 (3 M). The pore size distribution was measured with a Porolux 500 (Porometer) following the method described by Trzaskus et al. (2015). Based on the measured pore size distribution, the exclusion limits (cutoffs) of the membranes were calculated as defined to retain 90% of a minimum particle size (relates to the cumulative distribution at 90%; in short D 90 ). To evaluate the pore size distribution, the width was determined at the 90th percentile (relates to the range from D 5 to D 95 ) eliminating measurement noise at lowest and highest pore sizes.

| Membrane surface charge
To determine the zeta potential of the inner membrane surface, a single hollow fiber (90 mm length) was potted in a PE tube (80 mm length and   analyzer (Anton Paar). The streaming potential of membranes was measured in a 5 mM KCl electrolyte solution at a pH of 7.2 (±0.1). The zeta potential was calculated from the streaming potential via the Fairbrother-Mastin equation (Fairbrother & Mastin, 1924).

| Cell broth zeta potential
The zeta potential of the crude cell broth was measured in triplicates using 1.5 ml samples filled in a folded capillary zeta cell using a the Zetasizer Nano ZS (Malvern Instruments). The cell culture medium was measured as dispersant with a refractive index (RI) of 1.33, based on refractometry measurements (RE40D Refractometer, Mettler Toledo). Assuming a very low Debye length relative to the size of the colloids in the broth, the Smoluchowski approximation was used to calculate the zeta potential based on the electrophoretic mobility (Swan & Furst, 2012), and each sample was measured 30 times at 25°C following the manufacturer's recommendations.

| Scanning electron microscopy
Native and fouled membranes were either cut manually or frozen in liquid nitrogen before being broken manually. In brief, membrane fractions were fixed with carbon conductive tapes and carbon paint (DAG-T-502, Ted Pella) on specimen mounts, and vacuum-dried at 30°C overnight. A 10 nm chromium layer was sputtered on the sample with a Quorum Q150T ES (Quorum). The cross-section and surface morphology of the membranes was obtained using a scanning electron microscope (SEM; JSM-6010LA, JOEL) at 5 kV.

| Membrane filtration setup and experiment
Suspension-adapted BHK-21 SUS cells (derived from adherent BHK-21 cells, kindly provided by Dr. Boris Hundt, IDT Biologika) were cultivated in serum-free basal growth medium (BGM) in a 2.5 L DasGip glass bioreactor connected to a DasGip DCU controller (Eppendorf).
Cells were infected with YFV-17D (kindly provided by Prof. Dr. Mathias Niedrig, Robert Koch Institute Berlin) at multiplicity of infection of 10 −1 based on the plaque assay as described below. All membranes were prewetted with deionized water (dH 2 O), subsequently gently drained and connected to an external recirculation loop with a peristaltic pump (Watson-Marlow 120U). The membranes were consecutively tested in TFF mode at a fixed shear rate (γ) of 2,000 s −1 .
Therefore, volumetric flow ratesV f (ml/min) were adjusted based on the cross-sectional areas of all fibers of each module: (1) where f n is the number of hollow fibers and r the inner fiber lumen radius of individual fibers (mm). The permeate pump was set to a permeate flux rate J of about 33 L/hr/m 2 describing the ratio of the permeate flow rateV p (L/hr) to the total filtration surface area A (m 2 ) of all fibers in one module:= The permeate was transferred back into the bioreactor.
where TMP is the transmembrane pressure (mbar) and η m the dynamic viscosity of the medium (0.69 mPa s at 37°C). With the inlet equal to the outlet pressure, the TMP corresponds to the pressure difference between the feed and permeate stream. The fouling capacity of each membrane was described by the specific permeate with V as maximum permeate volume (L).
Samples of the bioreactor vessel and the permeate line were taken regularly, centrifuged at 2,000×g for 3 min and optionally stored at

| Virus quantification
Infectious YFV titers were quantified by plaque assay using stable porcine (PS) cells as described previously

| Rejection coefficient
The rejection coefficient σ reject was introduced to describe the fraction of product retained by the membrane and calculated as where C p is the YFV (PFU/ml), DNA or protein concentration (μg/ml) in the permeate flow, and C v (PFU/ml or μg/ml) the respective concentration in the bioreactor vessel.

| Structural and physicochemical membrane properties
The fiber wall thickness of most hollow fiber membranes was in a range between 0.10 and 0.15 mm, whereas large-pore PS (#6) and PE (#8) membranes were significantly thicker with 0.45 and 2.75 mm (Table 1).
Pore size distributions of membranes were determined by capillary flow porometry. The membrane-specific exclusion limit (cutoff) of the cumulative pore size distribution at 90% (relates to D 90 ) ranged from 0.08 μm to 1.69 μm (Table 2). Interestingly, the measured pore sizes differed from manufacturer's specifications.
Compared with nominal pore sizes (cannot be provided due to confidentiality agreements) four membranes had a larger effective cutoff by factors between 0.2 and 5.9, and four a smaller effective cutoff by factors of 0.1-0.8 (not shown here).
The width of the pore size distribution was described with the pore size width at 90th percentile and expressed in relation to the measured cutoff. Large-pore PE membranes tended to have a broader pore size width, while the PS membranes had very distinct pore sizes (Table 2; Figure S1).
Next, all membranes were examined with SEM imaging to investigate structural details. The material roughness was mainly assessed based on the frontal view of the inner membrane (Figures 2   and S2). While a highly jagged material surface was found for the large-pore mPES fiber, the roughness decreased from ME, PES (0.18 μm), PS (0.08 μm) materials to very smooth PS (0.34 μm), and PE structures. In addition, the porosity of the inner surface was qualitatively evaluated. SEM imaging revealed a remarkably high surface porosity for the PS (0.34 μm) membrane, which decreased from ME, the two PES, and the PE to the PS (0.08 μm) membrane. Due to the highly jagged material, visual evaluation of the mPES membrane was difficult. However, funnel-shaped pores were present, as equally observed for the PE membrane turning both membranes potentially susceptible for rapid particle entrapment.
Subsequent cross-section and frontal SEM imaging of the inner membrane helped to characterize surface roughness and overall porosity ( Figures S3 and S4). The mPES membrane had a very high surface roughness with distinct and deep valleys. The PES (0.18 μm) and ME membranes, whereas, had a flatter inner surface structure than the PS (0.34 μm), and PE membranes revealing a wavy surface.
The mPES material had a high porosity, followed by decreasing porosities with the large-pore PS, ME, PES, PE, and finally small-pore PS membranes. In particular, the front view of the outer surface revealed a strong asymmetric structure for most membranes except for the PE membrane ( Figure S4). A closer examination of the largepore PS membrane revealed a high overall porosity in the first inner half, which then became more compact to the outer side ( Figure S5).
Finally, the electrokinetic potential of membranes and potential foulants was assessed. First, the streaming potential of each membrane material was measured at pH 7.2 and 5 mM KCl solution to calculate the zeta potential. The zeta potential was about −24 mV for most materials, whereas the mPES material showed a slightly lower surface charge with −19.7 mV (Table S1). Then, the zeta potential of the culture broth containing infected cells (with extracellular vesicles, virions, and debris) was calculated. Based on the electrophoretic T A B L E 2 Overview on measured cutoff and pore size width (indicates pore size distribution) of hollow fiber membranes

# Material
Cut-off (μm; D 90 ) a Pore size width (μm) b During the filtration experiment, the membrane resistance increased fast for the two mPES, both PES and the ME membranes with only short periods of slower resistance development (Figure 3).
At maximum technical resistance, the permeate flow dropped and the silicone tubing on the permeate side collapsed due to low pressure at permeate side. Thereby, a V p of around 9-18 L/m 2 until termination was reached for most membranes (Table 3). For the tested PS membranes (0.08 and 0.34 μm), it took significantly longer before the maximum resistance was achieved resulting in permeate volumes of 30 L/m 2 and 75 L/m 2 , respectively.
F I G U R E 2 SEM images of the inner membrane surface of different hollow fiber materials. At given scale, a ×2,000-fold magnification allowed direct comparison of roughness, surface structure, and porosity of unused membranes. ME, mixed ester; mPES, modified polyethersulfone; PE, polyethylene; PES, polyethersulfone; PS, polysulfone; SEM, scanning electron microscope Subsequently, a selection of blocked membranes was subjected to SEM imaging. In particular, the mPES, the PES, and the ME membranes exhibited strong filter cake formation. Interestingly, the large-pore PS membrane did neither show surface-related deposition nor indications of pore blockage and cake layer formation ( Figure S6).

| Virus retention during TFF
While membranes were challenged, samples from the bioreactor broth and permeate were routinely taken and analyzed for infectious virus titer as well as DNA and protein concentrations. In the early filtration phase of small-pore membranes, virus titers in the permeate were already significantly reduced compared with the bioreactor vessel (9.0 × 10 4 PFU/ml; Figure 4). The small-pore mPES (0.09 μm) and PS (0.08 μm) membranes retained more than 99% of the infectious virus material, whereas almost 90% of the infectious material was retained by mid-pore PES (0.18 μm) and ME membranes (

| DNA and protein rejection
The rejection of DNA and protein contaminants was calculated based on depletion levels from the supernatant of infected BHK-21 SUS cells growing in BGM medium compared with the permeate (Figure 5). showed high rejection rates in the beginning, which then stabilized with a rejection coefficient of about 10% ( Figure 5).

| DISCUSSION
From a wide range of hollow fiber modules developed for various TFF applications (e.g., bioreactor perfusion, concentration, diafiltration, and clarification), eight commercially available membranes were selected and characterized for virus retention, DNA and protein contamination removal, and filter fouling. If available, a small-and large-pore retention. While small-pore membranes can be suitable to accumulate the product in the bioreactor, large-pore membranes can potentially be employed to continuously harvest virions (for YFV~50 nm). In both cases, it is desired to keep filter fouling to a minimum as it terminates the filtration process, and potentially ends in a complete product loss.

| Impact of general and physicochemical membrane properties on membrane fouling
To evaluate the impact of general and physicochemical membrane properties on membrane fouling and membrane blockage, observations were classified to predict their potential impact on membrane fouling. A high fiber thickness, a narrow pore size distribution and a high repulsion of foulants are considered to reduce filter fouling, while large pores allow general virus permeability (Table 4).
An increased fiber wall thickness is generally assumed to decrease permeate fluxes per driving force. In addition, a large contact surface allows for the adsorption of colloids, whereas intramembranous fluxes are increased in porous membranes that reduce membrane fiber blockage. The pore size distribution of membranes can be controlled to a certain extent by the manufacturing process, but is typically characteristic for the used material (Zeman & Zydney, 2017). For the PES, PS, and ME membranes, the 90th percentile of all pores was in a distinct range of about 25% in relation to the cutoff. The large-pore mPES and PE membranes, however, spread above 47%. Heterogeneous pore distributions are considered more susceptible to fouling as significant variation in filtrate flux along the length of the module occur, which turns large pores with higher local fluxes prone for concentration polarization and deposition until pore blockage. Thus, narrow pore size distributions have a uniform flux distribution and are generally considered better suited for long-term filtration operation (Jonsson, 1985;Table 4). The zeta potential was determined to assess repulsion effects.
Thus, an advantageous repulsive effect for all tested membranes  (Table 4). It should be noted that the zeta potential of the membranes was measured at lower salt concentrations, while the filtration was done under broth conditions with high salinity. It can be assumed that the streaming potential during filtration was lower than measured but values appear to correlate to the zeta potential of the culture broth (Breite et al., 2016;Schäfer, Pihlajamäki, Fane, Waite, & Nyström, 2004).

| Impact of membrane structures on membrane fouling
SEM imaging revealed significant structural differences of tested membrane materials, and properties can be equally assessed regarding their potential fouling behavior (Table 5). While a high roughness of the inner membrane surface can hinder direct pore blocking (steric exclusion of particles and nonflush deposition on highly fissured surfaces), a reduced overflow velocity in valley-like structures can equally enhance deposition (Marshall, Munro, & Trägårdh, 1993). Such loose deposits are particularly sensitive for cake compression, when negative pressure on the permeate side increases (Vrijenhoek, Hong, & Elimelech, 2001). This could be assumed especially for mPES membranes, which additionally possess a high specific surface area that potentially enhances particle adsorption. Deep valley-like pore channels, as observed for the PE membrane, and narrowed pores are also unfavorable due to enhanced particle entrapment and membrane blockage. In contrast, the PS membrane (0.34 μm) has a very smooth material and open pore structure, as well as a high overall porosity so that foulants can freely penetrate the membrane, but are finally retained in deeper, more dense layers. This can enable high initial fluxes, but as deposits enrich within the membrane and physical countermeasures (e.g., increased flow velocity and backflushing) may not allow to overcome corresponding problems, full blockage will be inevitable.
A size-selective and flat membrane surface, as observed especially for small-pore PS, but also for PES and ME membranes, enables thin boundary layers and optimum abrasive effects of the surface velocity (Choi, Zhang, Dionysiou, Oerther, & Sorial, 2005). This reduces concentration polarization (tendency for accumulation of foulants).
However, if the surface porosity is low, such filters can react sensitive on pore narrowing with increasing membrane resistance. In dependence on the pore size and the size of foulants, small-pore membranes (in the range of ultrafiltration application) may be even less affected by fouling due to steric exclusion for pore narrowing or pore blocking (i.e., 0.08 μm PS membrane).  (Cho, Amy, & Pellegrino, 2000;Cornelissen, 1997).
The increasing TMP (data not shown) compresses the filter cake, leading to full membrane blockage (Rana & Matsuura, 2010). A similar fouling tendency was observed for the small-pore PES membrane with low porosity. High permeate fluxes narrow scattered pores on the surface causing a quick reduction in the pore size, filter cake compression, and full blockage (Trzaskus et al., 2015). The short plateau in the development of membrane resistance for large-pore PES and ME membranes is, most likely, due to an equilibrium between deposition and foulant removal by overflow velocity until deposition dominates and the flux finally collapses. The PE membrane blocks potentially due to pore constriction and substantial pore closure. Interestingly, the PS membranes block only at notably high specific permeate volumes making them a candidate for longterm filtration operation. The fouling progression indicates an initial pore narrowing for the small-pore membrane, and an extended equilibration phase between deposition and foulant removal. The large-pore PS membrane with high porosity seems to be hardly affected by initial foulant-membrane adsorption and pore narrowing.
Its relatively high membrane thickness (approximately four times larger than other membranes) did not noticeably contribute to lower intrinsic permeability. Instead, it seems to provide a larger effective separation surface area contributing to a better resistance against overall filter fouling. The round-shaped material structure enables high fluxes across and within the membrane and mitigates adhesion of foulant particles. However, due to its asymmetric membrane structure and pore narrowing, an irreversible particle deposition in deeper layers of the membrane can eventually not be avoided (Henry & Brant, 2012;F. Wang & Tarabara, 2008). This is in agreement with SEM imaging of the blocked membrane. While a strong cake is formed on fast fouling membranes such as mPES, PES, and ME, the large-pore PS membrane does not exhibit any obvious foulants on the surface ( Figure S6, note that specimens were dried for observation, so that actual height of the cake layer could even have been greater during filtration operation). Therefore, foulants may be expected to be present at high quantities in deeper membrane structures. Notably, the observed membrane fouling progression is in close agreement with findings obtained for microfiltration processes (Trzaskus et al., 2015;Xiao, Shen, & Huang, 2013). It should be noted that the fouling of membranes is strongly linked to the material and process conditions tested. In accordance to the intrinsic membrane permeability (e.g., pore sizes, density, and physicochemical properties), optimal flow velocities and permeate fluxes can vary to achieve homogeneous fluxes along and through the membrane. Suboptimal conditions can otherwise favor local deposition and accelerate the progress of fouling.
Having understood fouling principles for the different membranes, product retention can be directly associated with membrane fouling dynamics. In the case of DNA and protein concentrations, here considered as impurities, their percentage rejection increased, possibly due to steric exclusion in narrowing pore channels, and increased repulsion from adsorbed foulants. Interestingly, the overall rejection was significantly higher for DNA than for proteins. Notably, the percentage rejection with PS (0.34 μm) and PE (1.68 μm) membranes showed a contrary trend. This observation may be explained by initial adsorption of DNA and protein to the membrane materials.
Once the adsorptive membrane capacity is reached, impurities may migrate unimpaired through the large-pore channels into the permeate (Cornelissen, 1997). Hence, the use of PS membranes can be equally important for related perfusion processes where expressed proteins are considered as product, but retained by PES membranes with a nominal cutoff of 0.2 μm (Karst, Serra, Villiger, Soos, & Morbidelli, 2016;Kelly et al., 2014). Alternatively, other studies identified the use of large-pore membranes of 2 μm and larger as a solution for production retention (Pinto, Napoli

| CONCLUSION
Our results highlight the importance of choosing the right membrane for intensified virus production and continuous product harvesting. We show that a selection based solely on nominal membrane pore size values reported by manufacuturers may not be sufficient. Instead, 3050 | membrane material and associated structural and physiochemical properties are decisive factors that determine filter fouling and eventually the "true" membrane pore size causing product retention. The widely used PES (0.18 μm measured cutoff) membrane fouled quickly, so that YFV titers but also protein concentrations decreased rapidly in the permeate flow. In contrast, the 0.34 μm PS membrane was highly permeable for YFV particles and enabled continuous product harvesting in small-scale hollow fiber modules and TFF mode. In this context, different process conditions (e.g., flow velocity and permeate flux) and filtration operations (e.g., hydrodynamic backflushing, inverting flow directions, and pulsed flow) can be investigated to improve performance of the PS-based perfusion processes even further.