Typical large-scale antibody recovery processes begin with harvesting Chinese hamster ovary (CHO) cell culture fluid (CCF) at the end of a culture cycle. The harvest involves separating secreted product (antibodies) from whole cells and cell debris present in the CCF. Traditionally, this is done using a tangential-flow filtration (TFF) system or centrifuge in conjunction with depth filtration (1, 2).

The existing CCF harvest process discussed here uses a disk-stack centrifuge in conjunction with a robust two-stage depth filtration train that ends with a sterilizing-grade 0.2-µm filter. Figure 1 shows a typical centrifuge harvest configuration. The depth filtration train removes cellular debris and potential carry-over cells from the centrate, in turn protecting the subsequent sterilizing-grade filter from fouling. In process transfer to a new manufacturing facility, we performed small-scale studies to evaluate several candidates for a single-stage depth filter train to accommodate existing equipment at the new facility.

Particle removal is achieved by depth filtration as the consequence of two separation mechanisms (3). The first is mechanical retention, in which particles are captured within the filter because of their physical size. They may be trapped on the surface of the filter, thus not entering the matrix at all, or captured while traveling through the tortuous path of the depth filter matrix (sieving) (4). Mechanical retention of particles causes a pressure increase across a filter as its matrix is blocked by accumulating particles. The second mechanism of depth-filtration is adsorption, by which particles that are smaller than the filter's sieving capability can be removed from the fluid stream (2, 5). This mechanism of action results from a net charge on the depth filter matrix.

Lack of adsorptive capacity in a depth filter may lead to particle breakthrough during filtration or after exhaustion of adsorptive capacity when saturation of the depth filter is reached. In addition, a depth filter can fail in its adsorptive capacity when flux through the filter reaches a shearing threshold (e.g., because of increased pore blockage and constriction), which can remove bound particles (3). The consequence of adsorptive capacity is thus very important for filters downstream of the depth filter(s) because they can experience a sudden onset of particle load and foul suddenly. In a harvest filtration train, that is typically reflected by the sudden loss of permeability in the sterilizing filter, which has a relatively low capacity for particulates. Therefore, it is important to monitor the pressure on both the depth filters and filters downstream from them during small-scale experimentation because both modes of failure (pressure build-up and breakthrough) can occur independently.

PALL LIFE SCIENCES (www.pall.com)

Materials and Methods



Our company's existing platform harvest depth filtration consists of loose-grade depth filtration media followed by tighter-grade depth filtration media that protect the final sterilizing filter. All depth filter media are lenticular format in 316L stainless steel housings. The filters we used in our study are described below. Table 1 lists the single-layer depth filter media with nominal pore ratings we used at large scale. Table 2 lists dual-layer, single-stage depth filters we tested in our scale-down studies. We also used small-scale capsule versions of the filters from Table 1 in the scale-down studies. Table 1: Single-layer depth filter media used at large scaleTable 2: Dual-layer depth filters used in scale-down studies

Depth filter media from Millipore Corporation (www.millipore.com) are composed of cellulose fiber combined with an inorganic filter aid (diatomaceous earth bound by polypropylene components), all encased in polypropylene structural components. The brand names of this vendor's single-layer and dual-layer depth filter media are Millistak+ DE and Millistak+ HC, respectively. The HC filter media are made up of two full-thickness depth filter layers, a coarse filter upstream and a fine one downstream, with an additional final layer of RW01 cellulose membrane. We used Millistak+ Mini filters in 60-mm disc format for our scale-down studies.

Depth filter media from Pall Life Sciences (www.pall.com) are composed of cellulose fiber, diatomaceous earth, and positively charged resin with polypropylene components. The brand names of this vendor's single-layer and dual-layer depth filter media are Supradisc and Supradisc HP, respectively. The HP filter media are made up of two full-thickness depth filter layers, a coarse filter upstream and a fine one downstream. We used SUPRAcap filters in 60-mm disc format for our scale-down studies.

Depth filter media from Cuno Inc. (a 3M company, www.cuno.com) are composed of cellulose fiber, diatomaceous earth, positively charged resin, perlite, and polypropylene components. We tested two brands — namely, Zeta Plus Maximizer and Zeta Plus Maximizer EXT for clarification. The former is made up of two half-thickness layers, whereas the latter is made up of two full-thickness layers and incorporates a higher charge density. We used Cuno Biocap filters in 60-mm disc format for our scale-down studies.

For sterilizing grade filters, we used Pall Fluorodyne EX dual-layer membrane filters. These consist of an upstream layer of hydrophilic polyethersulfone (PES) over a downstream layer of hydrophilic polyvinylidene fluoride (PVDF). Both layers have a nominal pore size of 0.2 µm.



DEFINITIONS

Throughput (L/m²) is defined as the volume filtered divided by the filtration area. This is independent of scale and is a constant factor used to determine small-scale flow rate and target volume through a filter.

Capacity factor (dimensionless) is defined as the achieved throughput at small scale divided by the target throughput. For example, an experiment that achieves a throughput of 200 L/m² in which the target throughput is 100 L/m² has achieved a 2× capacity factor. Any capacity factor exceeding 1× is often referred to as a safety factor.

Flux (L/m²/h) is defined as the flow rate (L/h) divided by the filtration area.

We obtained material for our scale-down studies from two different large-scale centrifuges, an Alfa-Laval BTPX-215 (www.alfalavalcentrifuge.com) and a Westfalia CSD-130 (www.westfalia-separator.com). Figure 2 shows our small-scale experimental set up. We used BD DTX Plus TNF-R pressure transducers (www.bdbiosciences.com) with a Netdaq network data acquisition unit Logger software from Fluke (http://us.fluke.com) to log the pressure data. Peristaltic pumps with appropriate silicon tubing gave the desired filter fluxes. In some experiments, we used a FilterTec CP-120 pump and data collection software from SciLog (www.scilog.com). Clarified CCF was collected in a vessel, and a balance provided real-time gravimetric data for volume and volumetric flow rate. We maintained constant flux (Pmax method) and verified it throughout each experiment using real-time gravimetric readings (5).

Figure 2 shows our scale-down model. We determined the target throughput for depth-filter sizing using the maximum expected volume of centrifuged CCF passing through the depth filter train at large scale and the available large-scale depth filter area. The large-scale flow rate was dictated by the operating flow rate of the centrifuge, and we derived the scale-down flow rate from the large-scale flux. The target volume at small scale was determined by multiplying target capacity by the small-scale filter area. We tested small-scale filter trains to at least 1.5× capacity based on a large-scale feed volume and various simulated large-scale flow rates. We chose the end-point pressure as a pressure drop of ~25 psid across any filter in the train.

Our scaled-down filtration train was designed by scaling target flux and target throughput as close to large-scale as possible. In some cases, discrepancies arose between the large-scale filter areas and relative filter areas of the small-scale devices. In such cases, target flux and throughput cannot be exactly matched for all filters in a train, so the filter with the highest flux at small scale was used as the basis for scale-down calculations. The rest of the filters in the train were operated at higher flux and higher throughput conditions to represent worst case. The filter flux or throughput targets tested at small scale always should be greater than or equal to the large scale values for each corresponding filter in the train (Table 3). As shown in Table 3, the DE50 filter operates under the highest flux at scale, so the scale-down flux is kept equivalent for it (using the DE50 filter as the basis for all scale-down calculations). Because of the available small-scale sizes, the remaining filters in the train were operated at a higher flux than the same filters at large scale. Table 3: Dual-stage filter train scale-down fluxes (LPM/m²) — large scale and small scale

Results and Discussion



Filter Screening: We screened depth filter media from Millipore, Pall, and Cuno to understand the relative performance of dual-layer depth filtration trains in terms of mechanical fouling and breakthrough. We compared their performance with a standard two-stage depth filtration train for harvest operations using centrate from the large-scale process.

We used centrate from the large-scale process to evaluate different single- and dual-stage depth filtration trains for harvest operations. Centrate of 70 NTU turbidity for this initial screening came from an Alfa Laval BTPX 215 disc-stack centrifuge operating at target conditions. Table 4 shows the filter trains tested and the resulting capacity and pressure-drop (d_P) endpoints (psid) corresponding to the capacity factor, which we calculated as the final volumetric throughput divided by the target throughput. The final pressure drops on the first stage, second stage (if applicable), and sterilizing filter stage correspond to d_P1, d_P2, and d_P3, respectively. Table 4: Filter train capacity and pressure drop (d_P) endpoints

When we compared the two dual-stage depth filter train controls, the L50 train exhibited breakthrough of both depth filters, whereas the tighter-grade L65 train provided more protection to the downstream sterilizing filter. We saw similar behavior for the tighter A1HC and looser B1HC filter trains. All Pall filters protected the sterilizing filter. The PDC1, PDD1, and PDE2 models provided the lowest pressure increases while protecting the downstream sterilizing filter.

The Cuno 120M10 train protected the downstream sterilizing filter, but the 120ZA10A train exhibited breakthrough. The latter was expected to have a higher adsorptive capacity that would protect against breakthrough; however, what we observed may be attributed to lower centrate residence time due to higher flux. The higher flux through that filter may also disrupt particle binding due to shearing velocity (3). Higher experimental flux through the 120ZA10A model may be attributed to the lower large-scale filter area (due to increased pad thickness), which would require testing at a higher flux relative to the 120M10 filter.

Based on those initial results, we chose Millipore A1HC and B1HC, Pall PDD1 and PDE2, and Cuno 120M10 filters for further evaluation. Although the PDC1 filter exhibited a lower pressure drop than the PDD1 filter did, we decided to select the top-performing filters with the largest retention ratings. And we added the Cuno 120M08 filter for evaluation to explore a looser-grade upstream layer.

Final Filter Selection: For final filter selection, we collected centrate samples from the new facility (four different full-scale runs with the Westfalia CSD-130 hydrohermetic centrifuge processed at small scale). The centrifuge was operated under various flow rates and bowl speeds to generate a range of turbidities and feedstock quality, which allowed us to explore the robustness of the various filter media.

Table 5 shows the filter trains tested with their resulting capacity and pressure-drop endpoints and feedstock attributes (% viability of cell culture and turbidity). We tested each run at the corresponding flux (e.g., equivalent centrifuge flow rate) from which the centrate sample had been collected. Table 6 summarizes filtration capacities for all the depth filter media tested at this stage. Table 5: Filter-selection feedstocks and endpoint summaryTable 6: Summary of small-scale filtration capacities at 25 psid endpoint relative to the required large-scale throughput

The PDE2 filter consistently exceeded our 1× capacity target with a comfortable safety factor for the entire range of centrate turbidities. We did not test the two-stage depth filter trains (controls) under all conditions because it was not an option in the new facility. We did not test the Millistak A1HC and B1HC filters under worst-case conditions for turbidity because they did not meet our target capacity at the median turbidity of 100 NTU. The Pall PDD1 filter also failed to achieve target capacity at 160 NTU, which represents the centrate obtained when the centrifuge was operated under worst-case conditions of flow rate and bowl speed. The Cuno 120M08 filter showed signs of mechanical fouling at ~0.7× capacity, which indicated that it probably would not reach the preferred endpoint of >1× capacity. (That run was terminated early due to tubing malfunction.) Because it has a tighter upstream layer than the Cuno 120M08 model, we tested the Cuno 120M10 filter with a different feedstock. Whereas that filter exhibited less mechanical fouling than the 120M08 filter, it did exhibit significant breakthrough to the sterilizing filters.

Our results indicate that a single-stage depth filter for harvest operations is feasible, with a comfortable safety factor across a range of feedstock quality. We chose the PDE2 depth filter media and further tested at laboratory scale for robustness and direct comparison with large-scale performance.

Filter Verification at Large Scale: We verified the robustness of our selected filter in conjunction with a direct comparison of large-scale performance for two runs. We used a range of centrifuge conditions to create a range of turbidities for small-scale testing, with a center point of 5,500 rpm and 60 L/min for most large-scale filtration. We set the other conditions to collect samples for small-scale testing; they represent less than 1% of the material filtered at large scale (so large-scale capacity represents 5,500 rpm and 60 L/min). Table 7 compares data from scale-down experiments with those obtained at large scale. The rows in bold represent a direct comparison, and all data are normalized to capacity. Table 7: Comparing PDE2 small-scale and large-scale endpoints

As shown, the PDE2 filter easily met 1.0× capacity at small scale for centrate collected at various centrifuge conditions. Figure 3 compares large-scale and small-scale PDE2 performance for each feedstock. Fouling trends and endpoints were comparable between scales. As Figure 4 shows, there was no significant increase in sterilizing filter pressure at either scale.

Fouling Characteristics: We used the median centrate turbidity value of 100 NTU for all fouling experiments. Quantitative evaluation of fouling trends is of limited practical value because of the complexity of depth filtration; however, simple qualitative analysis can offer some insight as to the mechanisms involved and may offer clues to future optimization. As Figure 5 shows, the general behavior of depth filter mechanical fouling is mostly linear, with some exceptions detailed below. This is typical of most depth and deep-bed filtration and indicates particulate deposition in the relatively open filter matrix during the initial stages of filtration (6,7,8). Each fouling profile is detailed below along with the corresponding sterilizing filter trend.

As Figure 6 shows, the DE50 filter in the standard two-stage train (DE50 → EKSP → sterilizing filter) exhibited no pressure increase throughout the filtration. The EKSP filter following it showed a generally linear fouling trend throughout the experiment. And the sterilizing filter began to foul at 1× capacity, which indicates exhaustion of the adsorptive capacity of one or both upstream filters.

As Figure 7 shows, the Millipore A1HC and B1HC depth filters both protected the sterilizing filter; however, they both exhibited significant mechanical fouling and reached the d_P endpoint of 25 psid before reaching 1× capacity. That fouling shows an initial nonlinear increase in pressure followed by a generally linear trend. The nonlinear segment may come from initial fouling of the RW01 membrane layers in the A1HC and B1HC filters (from particle breakthrough) before depth-filter ripening has occurred (6, 7, 9).

One interesting observation in the B1HC trend is the uneven pressure increase later in the filtration. Such behavior has been predicted under certain theoretical conditions in which a critical interstitial fluid velocity within the depth filter can be reached (because of increasing pore constriction), thus altering the filter's capture efficiency. Also, as retained particle clusters reach a certain size relative to the filter's interstitial velocity, they may break away and redeposit in the filter, causing a measurable redistribution of flow within it (8).

As Figure 8 (TOP) shows, the PDD1 depth filter protected the sterile filter from fouling and had mostly linear fouling characteristics. But the PDE2 depth filter did not protect the sterilizing-grade filter beyond 1.2× capacity (Figure 8, BOTTOM). That can be attributed to a looser-grade second layer in the latter, which provides less mechanical sieving of small particles. Although that depth filter may seem less desirable at this particular turbidity level, it is important to consider that over a range of turbidities it is more robust than others (Table 5).

As Figure 9 shows, the Cuno 120M10 filter's fouling curve is uneven, indicating significant changes in flow distribution throughout the run. It appears that flow redistribution may have occurred more than once throughout the filtration. Our analysis of flow rates during that filtration shows that flux remained consistent throughout the experiment, which indicates that the pressure perturbations indeed come from flow redistributions within the filter. As described above, such behavior might be explained as particle cluster release and reentrainment, which causes permeability changes within the filter during the run (8).

A Robust Alternative



We explored several dual-layer, single-stage depth filter media to find an alternative to a traditional two-stage depth filtration train. After carefully evaluating the options, we chose a robust filter and implemented it at commercial scale. The Pall PDE2 dual-layer depth filter was the most robust option, providing between 1.3× and >1.5× of the required capacity at all turbidities tested. We found it to be the most economically viable solution, providing significant cost savings relative to piping reconfiguration, new equipment purchase, and new equipment validation.

As observed from the filter fouling trends, failures of the depth filters we tested were primarily caused by a monotonic increase in pressure drop on the depth filter itself — or in some cases failure of the depth filter to protect the sterilizing-grade filter downstream. At all turbidities, an exponential pressure increase was rarely seen in the depth filter, indicating that even at our 25-psid filtration endpoint, no significant plugging of the internal pore structure of the depth filters had yet occurred. This general behavior may allow extrapolation of depth filter throughput based on shorter experiments; however, apart from mechanical fouling, depth filter breakthrough cannot be predicted and must be tested explicitly.

Interesting irregularities in the pressure profiles of the depth filters may be partially explained by theoretical predictions. Captured particles may form clusters in a filter matrix, break away after a critical fouling point, and redeposit themselves within the filter, causing changes in depth filter permeability throughout the filtration. These models also show that breakthrough failure may not be necessarily attributable to exhaustion of adsorptive capacity. Instead, such behavior could be attributed to reaching a critical interstitial velocity within the filter matrix (because of increased pressure drop across the filter caused by matrix plugging), which severely limits filtration efficiency and causes breakthrough of particulates (3, 8, 10, 11). One key to increasing future depth filter capacity may thus be to control particulate clustering (e.g., with additives) or modulate flow rate to avoid or exploit critical shear thresholds within a depth filter.