Evaluate retention against its actual task

A retention device in perfusion culture usually keeps viable production cells within the system while allowing liquid exchange. For a secreted protein, sufficient product must also reach the harvest stream. These are separate objectives. Almost cell-free filtrate can contain little product; improved product passage can coincide with greater passage of particular impurities. Selection therefore begins with three distinct requirements: which cells must remain, which product form should pass, and what composition may the next processing step receive?

The application considered here is a suspended mammalian cell culture producing an extracellular, dissolved product. The discussion does not transfer unchanged to viruses, cell therapies, aggregates or fixed beds. Where the cells are the product, the retained stream may be the desired harvest. Where ultrafiltration deliberately concentrates a protein in the reactor, low protein passage is intentional. Good retention and a good filter have no useful meaning without a product objective.

A complete process brief states working volume, cell concentration and size distribution, viability method, liquid throughput, intended duration, product assay, acceptable liquid inventory outside the reactor and the response to a necessary device change. It also requires biological comparison data for the pump, tubing and membrane. Comparing membrane area alone cannot compensate for an unsuitable circulation geometry. The calculations below provide diagnostic quantities; they do not authorise operation of a particular device.

Separate the loop, membrane and return flow

In conventional tangential flow filtration, TFF, culture moves along the membrane surface. Some liquid crosses the membrane as permeate; the remainder returns as retentate. Recirculation can greatly exceed the actual medium exchange. If 20 L/min circulates but only 2 L/h leaves the system, the harvest flow is 2 L/h. Internal circulation does not represent additional fresh medium and must not be counted twice as an external inlet and outlet in the liquid balance.

Alternating tangential flow, ATF, denotes periodically reversing flow along the membrane. A widely used implementation employs a diaphragm pump. Reversal of tangential flow does not mean that the harvest pump reverses its net delivery in the same rhythm. Pressure, time-dependent circulation and net permeate volume are different signals. A time average can conceal local loading peaks. Interpretation depends on where and how the specific system measures its variables.

The original study by Wang and colleagues on pump contributions in ATF and TFF demonstrates that pump selection substantially affected observed differences in the investigated arrangement. It does not establish a universal ranking of all ATF and TFF devices. A controlled comparison is more informative than an acronym: the same culture, an appropriate membrane, a traceable exchange rate and a documented exposure history.

Settling, acoustic and other retention concepts address the same separation task through different mechanisms. Membrane TMP is consequently not an appropriate measure for all of them. Each needs its own separation performance, return conditions, residence distribution and failure criteria. The distinction between suspension, carrier and fixed-bed cultures helps avoid equating a suitable cell environment with a particular external retention technology.

Four quantities that must not be confused

Permeate flow Fh may be expressed in L/h. Area-specific flux J = Fh/A is expressed in L/(m²·h). Effective membrane area A must be the area actually installed and operated. Equal flow distribution across parallel modules is a separate assumption; an unchanged total flow does not prove equal module loading. Flux measures neither cell retention nor product passage.

Observed product sieving is sP = ch/cr, the ratio of harvest to reactor concentration for the same analytically defined product form. It is dimensionless. At sP = 0.8, a litre of harvest contains 80% of the product concentration in a litre of reactor liquid under the conditions assessed. This does not mean that 80% of all product formed has already been harvested. Such a cumulative statement requires every relevant outlet and inventory change.

Similarly, observed cell passage can be defined as sX = Xh/Xr, with cell retention RX = 1 − sX. Numerator and denominator must describe the same cell class. A ratio of total cells in filtrate to viable cells in the reactor is not a cleanly defined retention measure. At very low filtrate concentrations, detection and quantification limits matter more than extra decimal places. A result below detection is not an exactly measured zero.

For a steadily flowing membrane channel, TMP = (pin + pout)/2 − pperm is a commonly used approximation. Three measuring positions do not reveal the complete local pressure distribution. All pressures require the same reference and, where relevant, a correction for elevation. Reversing flow requires an explicit definition of measuring positions, signs and cycle averaging. Axial pressure drop pin − pout is distinct from transmembrane pressure.

Four distinct retention-unit quantities
QuantityDefinitionWhat it does not establish
Liquid flux80 L/d ÷ 24 h/d ÷ 0.5 m² = 6.667 L/(m²·h)Product passage
Cell retention1 − 10,000/40,000,000 = 99.975%Exactly zero cell loss
Product sievingc permeate / c reactorNewly formed product mass
TMPSynthetic signal here: 0.08 → 0.30 barUnique fouling cause

Match paired samples and measurement times

A defensible sieving measurement uses a paired reactor and permeate sample. Sampling time matters, rather than the later time at which the laboratory releases the result. Permeate takes a finite time to travel from the membrane to the sampling position. With 0.12 L of tubing inventory and a flow of 0.04 L/min, the calculated mean transport delay is three minutes. This shift cannot be ignored during a rapid titre change. A vessel containing previously collected harvest is also a mixture of material from different times, rather than an instantaneous permeate sample.

Sampling must match the balance boundary. A reactor sample does not necessarily describe all material in a large external loop. With substantial concentration polarisation, concentration immediately next to the membrane differs from the mixed reactor sample. Sieving calculated from two accessible sampling points is therefore an observed system property. It is neither a pure membrane material constant nor a direct measurement of the original pore size.

Assay, dilution and sample handling must be consistent. If particles are removed before a protein assay, preparation at both locations should be comparable. A total-protein method may include medium components, whereas a product-specific assay measures a defined molecular class. Filtering or centrifuging a sample may remove bound product; an apparent concentration difference then partly reflects preparation. Sampling dead volume and residence time and correlated measurement uncertainties are directly relevant.

An apparent ratio above one first calls for examination of timing, assay, units and local concentrations. It should neither be clipped automatically to one nor dismissed immediately as a physical impossibility. Non-simultaneous samples, measurement variability or an unsuitable boundary can produce the observed ratio. Document the cause before interpreting the result as a filter property.

Worked case: unchanged flux, declining product sieving

The shared synthetic perfusion case represents 100 L of completely mixed total volume, including a loop assumed to mix completely with the reactor. Harvest flow is 80 L/d and membrane area is 0.5 m². A day contains 24 hours. Consequently, J = 80/(24 × 0.5) = 6.667 L/(m²·h). The six daily intervals use product sieving values of 0.95, 0.90, 0.80, 0.60, 0.40 and 0.25. These are chosen teaching inputs, not measurements from a product or filter.

At the same time, TMP, provided as a synthetic accompanying signal, increases from 0.08 to 0.30 bar. Constant flux shows that the specified net liquid delivery is still achieved. It does not demonstrate unchanged membrane operation. Declining sieving shows reduced product passage relative to reactor concentration. Rising TMP is consistent with additional hydraulic resistance, but does not uniquely identify its cause. A real investigation must distinguish viscosity, temperature, changed flow distribution, deposition and measurement error.

A separate hypothetical cell sample contains 40 million viable cells/mL in the reactor and 10,000 viable cells/mL in permeate. This gives sX = 0.00025 and RX = 99.975%. Nevertheless, with an 80 L/d harvest, 8 × 10⁸ viable cells leave through harvest each day. A high percentage retention does not imply an exactly loss-free cell balance. Biological relevance depends on inventory, growth, bleed and the process objective.

The dataset contains no experimental proof of a particular fouling mechanism. It deliberately requires liquid delivery, cell preservation and product harvest to be assessed separately. The exact product balance and collected mass are developed in perfusion balances. The interval table in CSV format and executable Python calculation support independent checking.

Test a cause with discriminating observations

The first diagnostic question is whether the change is real. Check the gravimetric balance, withdrawal record, paired-sample timestamps and raw pressure values. A kinked permeate line represents a different engineering problem from a loaded membrane. Increased permeate pressure can change TMP calculated from three measuring points even when retentate-side pressure remains unchanged. Sensor location relative to a valve therefore belongs in the diagnostic record.

The second question separates reversible concentration polarisation from persistent loading, to the extent that an authorised experimental arrangement allows this distinction. A controlled comparison with changed flux or adjusted recirculation can provide evidence. Such work must remain within the limits appropriate to the device and process; aggressively increasing pressure or pump output is not a general remedy. The intervention should also avoid changing the culture so much that the comparison concerns two different biological states.

The third question concerns the source of deposited material. Cell debris, released macromolecules, product aggregates and medium components may contribute differently. High viability among the intact cells still counted does not exclude previous cell lysis. A suitable supplementary lysis marker or particle analysis can therefore be more informative than another measurement of the same viability parameter. Appropriate methods depend on cell type and assay.

Before each intervention, document the prediction: if cause X is responsible, variable Y should respond under otherwise comparable conditions. Absence of that response weakens the hypothesis. Simultaneously replacing the pump, medium, temperature programme and membrane may improve operation, but leaves the cause unresolved. This distinction matters for subsequent scale-up and equipment selection.

Scale-up needs more than equal area per litre

Scaling membrane area at constant flux initially preserves only area-specific liquid loading. It does not automatically preserve shear exposure, concentration history, fouling or product passage. A larger loop can change pump passages per cell, tubing inventory and time spent outside the controlled reactor. Parallel modules introduce manifolds, different hydraulic resistances and possible maldistribution.

A transfer description needs at least two levels. At membrane level, specify area, channel geometry, material, separation characteristics, flux and relevant pressures. At system level, specify loop volume, pumping principle, flow profile, return arrangement, thermal and gas conditions, and access during a fault. The comparison should represent the actual culture exposure over the proposed duration, rather than just a brief initial water run.

Medium affects both hydraulic and biological behaviour. A water test may check a connection or sensor signal, but does not establish product passage in a protein- and particle-containing culture. A short cell-culture experiment does not automatically cover several weeks of loading. Accelerated challenge tests are useful when acceleration preserves the relevant mechanism; otherwise they create a different failure scenario.

Loop inventory also affects quantity planning. If 8 L of culture in an external loop is additional to 100 L in the reactor, total living-cell and product inventory occupies 108 L. If the 8 L was instead taken from a reactor initially filled to 100 L, total volume initially remains 100 L. These arrangements differ even though the nominal vessel is unchanged. The physical configuration must make the volume basis unambiguous.

Plan device changes and harvest handover before operation

A retention system may limit planned culture duration even when the reactor itself could continue operating. Decisions to stop a run or change a device should therefore use predefined process evidence: liquid delivery, product passage, cell loss, observed loading, quality and available interventions. A universal TMP limit for all systems would be inappropriate. The applicable equipment documentation provides the mechanical limit; the process evidence establishes quality and performance limits.

A device change requires an inventory balance, including possible dilution and time without normal exchange. If an old module retains 1.5 L at 1.2 g/L, it initially contains 1.8 g of product. A documented route must state whether that material is recovered, discarded or assessed differently. It must not appear simultaneously as harvest and loss. Flush liquid also needs an unambiguous destination.

The next processing stage needs the time-resolved harvest flow and composition. Declining sieving may reduce harvest concentration while reactor titre rises. Capture then receives a different liquid-volume-to-product ratio. If the enriched reactor inventory is processed later, a concentrated terminal harvest creates an additional load. Both streams should be incorporated into the harvest campaign plan.

A useful final report consequently distinguishes material disposition, operating duration, analytically observed condition and the reason for stopping. Completing the programmed duration is not by itself proof of successful product recovery. Conversely, deliberately ending earlier can, in a specific process, deliver more acceptable product per occupied equipment time than extending the culture as long as mechanically possible.

Diagnose unequal loading of parallel modules

Two parallel modules of 0.25 m² each provide 0.5 m² in total. At 80 L/d combined harvest, apparent average flux remains 6.667 L/(m²·h). If the split is 60 and 20 L/d, however, individual modules operate at 10 and 3.333 L/(m²·h). The overall average conceals a threefold difference in area loading. A measurement only after the combined collection point cannot determine that split.

Product passage behaves similarly. If individual permeates contain 0.5 and 1.0 g/L, the daily mixture carries 60 × 0.5 + 20 × 1.0 = 50 g in 80 L, or 0.625 g/L. The arithmetic concentration mean would be 0.75 g/L and would be wrong. Actual volume or mass flows determine mixture weighting. If the split changes over time, that weighting also needs integration.

Discriminating diagnosis therefore considers separate module flows, pressure losses and corresponding samples where the actual system permits. A combined signal remains useful for the external balance but cannot uniquely locate a module deviation. Flow distribution particularly matters when scale is increased by adding parallel modules. Equal total installed area does not prove that individual surfaces perform the same task under the same conditions.

An unusual sieving ratio is initially an observation

A measured permeate-to-reactor concentration ratio may apparently exceed one. This does not automatically imply a filter with negative retention. First check whether both samples represent the same product form and time state. When reactor titre falls rapidly, a delayed permeate sample may still contain product from the earlier higher state. A collection vessel can amplify this timing effect.

Release of previously bound product can also create a transient balance different from steady passage. An additional surface inventory and its change then belong in the model. The equation sP = ch/cr remains a definition of the observed ratio, but does not by itself describe an unchanging membrane property. Without further observations, the ratio proves neither enlarged pores nor a particular binding mechanism.

A practical check compares raw concentrations, dilution factors, sample locations, collection times and analytical methods. It then investigates possible transport delays and stored product. The objective is a balance explaining all observed quantities with consistent assumptions. An unusual value should neither be discarded automatically as an error nor promoted as unexpected performance. It marks a point where the current model or data capture is insufficient.

Account for residual quantities during a module change

An external loop can contain significant liquid and product inventory. Suppose an old module and its isolated lines retain 2 L at 1.5 g/L. Discarding this liquid outside the harvest removes 3 g of the defined product. A new module filled with 2 L of product-free liquid may restore volume but cannot restore the lost product. An unchanged reactor-level reading would not reveal the event.

If the original 100 L system boundary already includes the loop, its content must not be added to initial inventory again. If the boundary changes during an analysis, explicitly transfer inventories at that transition. The same principle applies between reactor, harvest collection and recovery. Otherwise an apparent mass jump may result from a changed definition rather than biology.

A complete change record therefore separates isolated, recovered and newly added quantities. It also records duration and conditions while normal exchange is interrupted. Whether a particular action is technically, hygienically and organisationally acceptable follows from the actual system and process concept. The example is not a module-changing procedure; it identifies the quantity information needed to analyse an acceptable change correctly afterwards.

Check your reasoning: which information is missing?

Question 1: A system reports 100 L of harvest on each of two days. Reactor titre is initially 1.0 and later 1.5 g/L. May a 50% increase in product harvest be reported? No. Harvest titre or integrated product mass is missing. With sieving values of 0.9 and 0.4, the corresponding instantaneous harvest concentrations would be 0.9 and 0.6 g/L. Without the time course, multiplying daily volume by a single sample remains an approximation.

Question 2: No viable cells are detected in filtrate; the detection limit is 500 cells/mL and the reactor value is 20 million cells/mL. What reporting-threshold calculation is possible? 500/20,000,000 = 0.000025 corresponds arithmetically to 99.9975% retention. This is not a statistically supported lower bound: a detection limit is not an upper confidence limit on the true filtrate concentration. Exactly 100% retention is not established either. Defensible bounds additionally require suitable sampling, method and uncertainty information. The VIM detection-limit definition explicitly considers false-negative probability.

Question 3: Two parallel modules provide a total area of 1 m² and deliver 10 L/h. Overall flux is 10 L/(m²·h). Does this prove that both modules operate at that flux? No. If one delivers 8 L/h and the other 2 L/h, each with 0.5 m², their individual fluxes are 16 and 4 L/(m²·h). Individual loading requires module flow measurements or a defensible demonstration of distribution.

These answers expose the same pattern: an attractive aggregate can conceal the distribution that matters. The next investigation should provide exactly the missing information. This may be a properly timed sample, an additional quantity measurement or separate module observations. Additional data are useful when they can change a specific decision.

Sources and next steps

Repligen: XCell ATF systems documents a specific commercial retention system. The page does not replace a manual applicable to the installed configuration. Performance promises and general superiority claims are not adopted here.

Liang and colleagues: CHO perfusion strategies, 2023 provides a useful counterexample to silently assuming complete protein passage: the investigated arrangement uses a 50 kDa membrane for concentration culture. Its reported reactor titre must therefore not be equated with continuous, complete protein removal.

Continue with liquid, cell and product balances and productivity on consistent time and quantity bases. The accompanying files contain an independently defined teaching case. Its coefficients were neither estimated from the linked experiments nor presented as properties of their equipment.

Terms in this chapter

32 terms explained in the glossary