Considerations for Scaling-up Depth Filtration of Harvested Cell Culture Fluid

Published on: 
, , , , , ,
BioPharm International, BioPharm International-03-01-2009, Volume 22, Issue 3

Data on the performance and variability of different formats.


The use of centrifuges to remove cells in combination with depth filters to remove colloids has become widespread. High colloid concentrations and allowance for process variability can lead to large single-use filter assemblies of several hundred square meters. This study measures the performance differences and variability between multilayer depth filters of different formats and sizes using a Chinese hamster ovary (CHO) centrate challenge.

Cell debris and colloids are removed from centrifuged harvested cell culture fluid (HCCF) by depth filters to extend the capacity of the downstream sterile filters. The successful efforts to improve factory productivity by increasing bioreactor titers often leads to higher cell densities, cell lysis, and higher colloid concentrations that can lower depth filter capacity. Allowance for process variability and scaling can require large, single-use assemblies of several hundred square meters.


Depth filters that show high capacities for this application are composed of cellulose fibers impregnated with diatomaceous earth and a polymeric amine binding resin. These depth filters are typically fed by the centrifuge and operated at a constant flow. As particle laden fluid is passed through these depth filters, the submicron colloids are removed and the depth filters show an increase in pressure drop across them. When the pressure drop reaches an operational limit, here taken as 15 psi, the filters are considered to have plugged. The amount of fluid that has passed through the filter is then considered to be the design capacity, expressed here as liters per square meter of filter area.


This study measures the performance differences and variability between multilayer depth filters of different formats and scales using a CHO centrate challenge. Variability from other sources was minimized to see the effect of scale and format more clearly. A single 12,000-L bioreactor production-scale batch of Chinese hamster ovary (CHO) cells expressing a monoclonal antibody product was used to challenge all the filters. Use of a single lot removes lot-to-lot harvest variability so that capacity differences between filters can be seen more readily. The centrifugation of the harvest took several hours, and therefore, controls were taken to assess if differences in hold time before centrifugation affected filter capacity.

Figure 1. Depth filters in a) Lenticular stack, b) Pod, c) Mini capsule formats

The different filter formats and sizes used in this study are shown in Figure 1 and Table 1. The Millistak+ A1HC filter used in this study contains multiple layers of cellulose filtration media and a cast membrane. A single manufacturing lot of filter cellulose media was used for this study. Use of a single lot removes lot-to-lot filter media variability so that capacity differences between filters of different formats and scales can be seen more readily.

Table 1. Test filter device formats and sizes

A parallel run between a 1.8 m2 lenticular stack device (Stack) and a 1.65 m2 Pod filter (Pod) assembly (one 1.1 m2 Pod device and one 0.55 m2 Pod device in parallel), allowed a more direct comparison of the different formats at comparable sizing. Large-area filters and assemblies were run in the manufacturing plant while small area filters were run in the pilot plant. Replicate 0.11 m2 Pod controls were run in both the pilot and manufacturing plants to assess any differences in performance between the two test locations.

Test operating conditions were the same for each filter device and size to minimize operating condition variability. Each filter was flushed with 50 L/m2 of water for injection (WFI) at a flux of 100 LMH (liters per m2 per hour). This was lowered from the recommended 600 LMH because of pump limitations. HCCF centrate was pumped at a constant flux of 111 LMH, representative of manufacturing-scale operation. Feed pressure and liters filtered were recorded using a data acquisition system during processing for small-area devices and recorded manually for large-scale devices. Filtrate quality was assessed by measuring turbidity every 40 L/m2. The run was stopped when the feed pressure reached 25 psid, past the 15 psid design limit. Pooled depth filter filtrates from each test were also filtered through a 0.2 μm sterile filter. These sterile filter tests were run at a constant 10 psi feed pressure and the filtered volume recorded over time. All depth filter formats were tested simultaneously and three replicates of the test plan were completed sequentially. Replicate testing was used to determine variability in results. There were 27 separate depth-filter tests as summarized in Figure 2.

Figure 2. Experimental design: there were 27 separate depth-filter tests


The pressure drop across each filter was calculated by subtracting the measured filtrate pressure from the feed pressure at each time point. This pressure drop was then divided by the constant filtrate flux of 111 LMH to calculate a hydraulic resistance as psid/LMH. Figure 3 shows the increase in filter resistance with loading for the 27 tests.

Figure 3. Depth filter resistance profiles

The filter resistance profiles showed a consistent shape for most filters, starting at around 0.025 psid/LMH and remaining fairly flat, then rising fairly rapidly. The initial variability is attributed to some air locking which disappears after a 0.05 psi/LMH resistance once the membrane intrusion pressure has been reached. The Mini capsule filters (Mini), however, tended to show increasing resistances at higher loading compared with other filters that tended to lie together.

Figure 4 shows the filtrate turbidity profiles for the 27 tests. These values are much lower than the average HCCF feed turbidity of 91 NTU showing that particles are being removed by the filters. The turbidity profiles for most filtrates showed the typical shape, a slow initial rise followed by a steady value. Contributing to the lower initial values is dilution by held-up WFI used for flushing at about 10 L/m2. Run 3 for the 0.45 m2 Stack was an outlier with higher turbidities from the outset.


Figure 4. Depth filtrate turbidity profiles

Filtrate quality was also assessed by running the pooled depth filter filtrate through a 0.2 μm sterilizing filter. Sterile filter plugging is assessed by calculating the ratio of the sterile filter hydraulic resistance at each time point divided by its initial hydraulic resistance for each test. This ratio is obtained by numerically differentiating the weight versus time data. Figure 5 shows the calculated ratio for 21 tests. Run 3 for the 0.45 m2 Stack was an outlier with rapidly increasing resistances with loading. This mirrors the higher turbidity results. The remaining resistance ratios show consistently low resistance increases. There are some sporadic spikes in the ratio attributed to accidental bumping of the scales.

Figure 5. Sterile filter resistance ratio profiles


Turbidity Outlier

The housing was disassembled for the 0.45 m2 Stack that was an outlier on turbidity and sterile filter profiles. The top gasket seal had failed to contact the adjacent seal as shown in Figure 6. The nonintegral seal is considered a source of feed bypass and the data from this run was not considered in further variability analysis.

Figure 6. Lenticular stack seal failure

Plugging Model Analysis

The resistance profiles for the different filter formats and sizes were fit to different plugging models (see Appendix for details): gradual pore plugging, cake filtration, intermediate plugging, complete pore blockage, cake-complete combined model, second-order polynomial, and third-order polynomial. Figure 7 shows an example of the model fits for one resistance profile. The empirical polynomial models fit the data best using more adjustable parameters. Of the mechanistic models, the combined cake-complete plugging model provided the best fit of the data. This is consistent for the multilayer filter used in this study that has a cast membrane layer in series with the cellulose media.

Figure 7. Plugging model analysis

Table 2 shows that the combined cake-complete plugging model provided a good fit for all devices, and the magnitudes of the best-fit parameters were comparable between formats. This provides a more quantitative assessment that the resistance curves are similar for all devices and suggests that all format types plugged by similar mechanisms with similar magnitudes of feed and filter interactions.

Table 2. Plugging model analysis

Filter Capacity

Filter capacity was assessed for each test by determining the L/m2 loading when the depth-filter resistance reached 0.15 psid/LMH (~15 psid). As shown in Figure 8, the depth-filter capacities were comparable and repeatable at about 400 L/m2 for all filter types with the exception of the Mini filters, which consistently showed 25% higher capacities of 500 L/m2. If the Mini filters are excluded, there is no statistical difference in capacity between the remaining devices. The Mini filter has more variability than the other devices.

Figure 8. Depth filter capacities at 0.15 psi/LMH resistance versus format and scale

Filtrate Quality

Figure 9 shows that the turbidities were comparable and repeatable at about 5 NTU with the exception of the Mini filters, which showed 15% lower turbidities at 4.25 NTU. The magnitude of the difference was small (<1 NTU) and its result on the downstream sterile filter capacity appears to have been negligible. It is likely this difference is because of the fluid composition averaging that occurs in slowly filling the vial for turbidity measurement using a Mini filter as opposed to somewhat instantaneous filling using larger sizes. Some effects could also be because of real differences in the devices or in measuring equipment or operator procedures. The 25% higher capacity of the Mini filters cannot be attributed to nonintegral devices by turbidity measurement.

Figure 9. Depth filter filtrate turbidities at 0.15 psi/LMH resistance quality versus format and size

Sources of Variability

Batch-to-batch variability in capacity at the manufacturing scale includes variability in the HCCF centrate feed, filter lot, and operation. For this study, we minimized variability in the HCCF centrate, limiting it to hold-time differences and labeled here as sequence. We restricted filter variability to different formats, sizes, and the variability within a single lot of filter media labeled as device. Variability in operation was limited to test location. The capacity data was analyzed by ANOVA using Minitab software. The sources of variability are summarized in Table 3.

Table 3. Sources of capacity variability

The ANOVA analysis confirms that the 23.5 cm2 Mini is a major source of capacity variability (Lutz, 2006). Excluding these devices causes the total capacity variance to drop 6x, from 4,072 (L/m2)2 to 735 (L/m2)2. This corresponds to a RSV = 6.8% (relative standard deviation = standard deviation/average). Note that the variances were assumed independent, so they are additive while the RSV are square roots of the variance and do not add. The variance distribution is shown in Figure 10. The variance is dominated by location with a RSV = 5.1%, which includes operator, instrumentation and experimental error. Sequence (capacity of first versus final runs) caused mean filter capacity to decrease approximately 10% with a RSV = 3.3%. Size and format are negligible contributors, indicating that there are minimal differences in filter capacity between the remaining laboratory-scale Pod filter (LSP), Pod, and Stack formats. Device variability with a RSV = 3.0%, is associated with filter variation in a single lot.

Figure 10. Variance components (excluding Mini devices)

It was anticipated that size would affect variability more significantly. Consider a larger filter or filter assembly with area A composed of n equal small filter elements of area A. If these small elements have a mean permeability μ and a standard deviation σ, the assembly of n elements has the same mean permeability μ and a standard deviation of either σ if the elements are perfectly correlated or σ/√n if the filter areas are statistically independent. The actual filter should lie between these extremes and show a reduction in variability with surface area. This can be called an averaging effect. This data shows that from the 270 cm2 LSP device upward, the variability is somewhat constant and any averaging effect is negligible compared with the other sources.

Excluding Mini devices, Figure 11 shows that the resistance profiles for all the remaining tests lie very close together. The capacities at 0.15 psi/LMH are 400 L/m2 +/–10%. LSP devices with 270 cm2 are the smallest-scale devices that are comparable with large- scale filter assemblies. It also suggests that the Stack and Pod formats have equivalent capacities and can be used interchangeably. The Mini is not a preferred scaling device.

Figure 11. Depth filter resistance profiles (excluding Mini devices)

Mini devices may perform differently as a result of having low area, in which the averaging effect has not yet reduced scale variability. It also may perform differently because the peripheral seal region is a larger fraction of the total filter area in the Mini compared with the other filter formats. The region near the seal may have different flow because of filter media compression. It has been observed in other testing (data not shown here) that the Mini capacity can vary up to +/–30% among devices or between the Mini and larger devices.

Manufacturing-Scale Variability

A review was done of 14 different 12,000-L manufacturing-scale runs of the same monoclonal antibody product. The way these runs were performed allowed the resistance to increase to 0.15 psi/LMH and the capacity determined. The batch-to-batch RSV was 10%. This includes batch-to-batch variability arising from feed, filter, and operating conditions. A comparison with the format study RSV = 6.8% confirms expectations that additional variability in capacity arises from differences in feed and filter media lots.

The manufacturing-scale capacities closely followed a normal distribution when graphed on a quartile plot. Using a safety factor of 1.4 sizes a filter assembly at 1.4 times the mean capacity, four relative standard deviations above the mean capacity. The chance of requiring a larger area is given by the area under the normal distribution as one in 10,000 batches.


All the filters appeared to plug by similar mechanisms. The LSP, Pod, and Stack depth filter formats can be applied to the clarification of harvested cell culture centrate with <±10% difference in capacity for sizes ranging from 270 cm2 up to 1.8 m2. These different filter formats also show comparable filtrate quality by NTU and Vmax filterability. No significant averaging effect of area was seen on variability. The variability among these devices was roughly constant.

The 270 cm2 LSP and larger devices are also recommended for accurate scaling. They are useful to assess filter capacity, filtrate quality, and expected variability at scale. The equivalent performance of the Pod and Stack formats indicate that they can be used interchangeably without affecting sizing or filtrate quality. A safety factor of 1.4 for sizing is expected to allow a filtration system to compensate for random batch-to-batch variations in feed, filters, and operating conditions. Other applications with larger variability in capacity because of centrifuge operation or harvest quality may require larger safety factors.

The 23.5 cm2 filters showed capacities that differed from the pod and lenticular formats by an average of 25%. The 23.5 cm2 filter device is useful for rough screening of different filter matrices, rough sizing (within +/- 50%), and filtrate quality evaluation.

This article does not represent an Amgen endorsement of any filters described herein and is not meant to imply that Amgen uses any of these filters for clinical or commercial manufacturing.


The authors would like to acknowledge the assistance of Karin Stephens from Amgen and Anne-Marie Braun, Matthew Daley, Jeffrey Shumway, and Ben Cacace from Millipore in the planning, execution, and analysis of this study. The significant assistance of Jonathan Royce, formerly with Millipore, is also acknowledged.


Table 4 summarizes different plugging models that describe how pressure varies with L/m2 throughput or normalized volume at constant flux (Badmington, Bolton, Hermans, Hermia, Zeman). The polynomial models are strictly empirical. The other models are associated with different mechanisms of filter plugging. These models can be fit to data to obtain values of the plugging parameters and a statistical measure of goodness of fit. The parameter values obtained have physical meaning based on the associated model.

Table 4. Constant flux plugging models

Herb Lutz is principal engineer and Mark Blanchard is principal research scientist, both at Millipore Corporation, Billerica, MA, 310.309.0984, Herb_Lutz@millipore.comIan Abbott is senior engineer, Ananth Parampalli is engineer, George Setiabudi is senior engineer, Vijay Chiruvolu is director, Megumi Noguchi is senior engineer, all at Amgen, Fremont, CA.


1. Badmington F, et al. Pharm Techn. 1995;64–76(1995).

2. Bolton, GR, et al, J Memb Sci. 2006;279:625–634.

3. Hermans PH, et al. J Soc Chem Ind. 1936;55T: 1–4.

4. Hermia J, Trans IChemE. 1982;60:183–187.

5. Lutz H. Recovery of biological products XIII Conference, poster;2006 Apr.

6. Zeman LJ, et al. Microfiltration and Ultrafiltration, Marcel Dekker, New York, NY;1996.