Perfusion, explained: three flows, one balance, and where the product goes

Research 38 primary sources read, 34 cited: papers, vendor-authored studies and ICH Q13 29 Sep 2026

22 to 32%

Perfusion keeps a cell culture running for weeks by feeding it fresh medium and taking spent medium out, while a device holds the cells back. It is the upstream half of most continuous biomanufacturing schemes, and ICH Q13 now uses a perfusion bioreactor as its worked example for therapeutic proteins ^S01. We read 38 primary sources, from 2011 to 2026, to explain how it works and where it costs you. The short version: a perfusion culture is a balance between three flows, and two of them deserve most of the attention on cost: the medium you pay for and the product you throw away with the bleed.

The six findings in short

• Perfusion is three flows in balance, and the bleed is the steering wheel. Fresh medium in, cell-free harvest out, and a bleed of culture that removes the cells the culture grows. At steady state the bleed is what holds the cell density on target. Confidence: high.
• The medium is the bill. The medium each cell needs per day (the CSPR) sets how much you buy. Published steady runs used between about 10 and 70 pL per cell per day, and in two studies getting below about 30 took reinforced or purpose-built media. Confidence: high.
• The bleed throws product away. In a seminal CHO study, bleeds carried off 22 to 32% of all antibody made. Confidence: medium, one detailed study plus supporting ones.
• The filter can hold the product back, and it gets worse with time. Membranes meant to stop only cells also held back part of the antibody, and one team saw retention climb from 20% to 60% over a run. Membrane structure and the pump make a large difference. Confidence: medium.
• Devices without a membrane let the product through but let cells slip. Acoustic settlers, inclined settlers, hydrocyclones and spiral channels kept roughly 84 to 99% of cells, and less as density or flow rose. Few of these data are from CHO. Confidence: medium to low.
• You do not need a full perfusion plant to benefit. Running only the seed step (N-1) in perfusion raised fed-batch titres 1.5 to 2 times at two companies. Confidence: high.

How we built this

We read 29 sources in full text and 8 from the abstract only, plus the ICH Q13 guideline. Many are written or co-authored by suppliers of the equipment or media they test, and we flag each one in the source list: Sartorius ^{S11 S13 S37} (and a co-author of ^S02), Merck Life Science ^S06, GE Healthcare, now Cytiva ^{S03 S04 S17 S33}, Cytiva ^S05, SonoSep ^S18 and Thermo Fisher ^S14. The two Clincke papers from 2013 ^{S03 S04} predate our usual window; we use them because they remain the most complete open data on CHO perfusion at very high density. Where a paper quotes a number from earlier work, we say so and do not attribute it to the citing paper. Each conclusion carries a confidence level: high means several independent sources agree; medium means one detailed study or one company’s data; low means little or no CHO data.

Fed-batch, perfusion and the three flows

A fed-batch culture starts with a few cells, adds concentrated feed and ends after about two weeks, when everything is harvested at once. Perfusion never stops feeding. Fresh medium enters continuously, a retention device lets clarified liquid leave as the harvest, and a separate bleed stream removes whole culture, cells included ^S11. Once the culture reaches its target density, the operator keeps it there by adjusting the bleed: the bleed is the manipulated variable ^S11.

The logic is simple bookkeeping. Every day the cells divide. If the device keeps every cell in the vessel, the only way to stop the density from rising is to remove, through the bleed, as many cells as grew. In the seminal KTH work, CHO cultures held at 20 to 35 million cells/mL with about 1.5 vessel volumes of medium a day grew at about 0.35 per day, and daily bleeds held them there ^S03. At pilot scale, a 30 L steady state at 100 million cells/mL ran at 1.5 vessel volumes a day, of which 0.1 to 0.3 went to the bleed and 1.2 to 1.4 to the harvest ^S05.

How dense can it get? The same seminal work reached 214 million cells/mL with TFF and 132 million with ATF, and held 90 to 130 million with bleeds for more than two weeks ^S03. Those are records, not operating points. More recent runs sit lower: about 54 million cells/mL at lab scale with an automated bleed ^S37, about 70 million at AstraZeneca ^S27, 85 to 100 million in a Merck & Co. run ^S21 and 100 million at 30 L pilot scale ^S05. An industry consortium uses 50 to 120 million cells/mL as its design basis for commercial N-stage perfusion ^S02; that is a planning assumption, not a measurement.

Try the balance yourself

The simulator below applies those three bookkeeping rules. Move the sliders and watch where the medium, the cells and the antibody go. It is an illustrative teaching model: it assumes a true steady state, constant growth and productivity and no nutrient limits, and it is not validated against a specific process. The ranges on the sliders follow the published values discussed in this article.

The medium is the bill

The key number for cost is the cell-specific perfusion rate, or CSPR: how much fresh medium each cell gets per day, usually in picolitres. Multiply it by the cell density and you get the medium flow. Hold CSPR constant and a denser culture needs proportionally more medium: in the KTH ATF run that reached 132 million cells/mL, the perfusion rate reached 6 vessel volumes a day ^S03.

So the whole game is to lower CSPR without starving the cells. The KTH cultures ran at about 50 to 70 pL/cell/day ^S03. In the KTH and Cytiva pilot work, the plain medium supported growth at 33 pL/cell/day but performance fell below that unless it was reinforced with concentrated feeds; with reinforcement the production bioreactor ran at 25 and then 15 pL/cell/day ^S05. BOKU and GE designed clone-specific media that brought the minimum down to 10 to 30 pL/cell/day while still reaching 200 million cells/mL ^S17. AstraZeneca describes about 20 pL/cell/day as typical for CHO in its proprietary media ^S27. Not every low value needed special media: a Merck & Co. run held 75 million cells/mL on 0.75 vessel volumes a day of a commercial basal medium, which works out to about 10 pL/cell/day (our calculation) ^S20.

More medium also buys less than you might expect. Boehringer Ingelheim found that in N-1 perfusion, 176% more medium gave only a 63% higher final cell density ^S07. Control helps: Bristol Myers Squibb cut medium use by about 25% by pacing the flow to the cell density measured online with a capacitance probe, instead of following a fixed schedule ^S08. Perfusion media are generally not sold off the shelf and have to be designed per process ^S17, although they can be built from fed-batch basal media mixed with concentrated feed ^S02.

The bleed throws product away

The bleed removes whole culture, and whole culture contains antibody. In KTH runs of more than 17 days at about 25 million cells/mL, the daily bleeds carried off 22% of all the antibody made with ATF and 32% with TFF; the harvest captured 71% and 52% ^S04. MacDonald and colleagues state that bleeds typically remove 10 to 30% of the culture volume each day, a figure they take from earlier work ^S14.

Two routes to cut that loss appear in the sources. One slows growth so less bleed is needed: in a four-day Merck Life Science scale-down model, growth-inhibiting additives raised the share of exchanged volume that went to the harvest from 66% to 80 to 82% in one cell line ^S06. The other engineers the cells: death-resistant CHO lines ran for weeks without any bleed at similar volumetric productivity ^S14. Both are early: one is a vendor study of its own supplements in spin tubes, the other a research cell line.

When the filter holds back the product

ATF and TFF both push culture along hollow fibres whose walls let liquid through and stop cells. ATF uses a diaphragm pump that pulls culture into the fibres and pushes it back, so the flow alternates; TFF uses a pump that drives it one way round a loop. The membranes are chosen to pass the antibody. In practice they often do not pass all of it.

With 0.2 µm polysulfone fibres, the KTH team measured 93 to 158 mg/L of antibody in the harvest while the bioreactor held 176 to 243 mg/L: the fibre kept part of the product inside ^S04. The retention was worse with TFF than with ATF ^S04. Boehringer Ingelheim traced TFF’s poorer sieving (the share of antibody that passes the membrane) to cells broken by the peristaltic pump, and found a low-shear centrifugal pump gave ATF-like results ^S28. WuXi Biologics, running a bispecific antibody through 0.2 µm PES fibres on ATF without a bleed, saw retention rise from 20% to 60% over the run ^S26.

The material and the structure of the membrane matter more than its nominal pore size. In one WuXi study, polysulfone fibres retained 15 to 43% more product than PES ^S30. At Merck & Co., a 0.65 µm membrane with narrow surface pores fell to 70% sieving or less within three weeks, while a 0.2 µm PVDF membrane with wide surface pores averaged 98% and stayed at or above 85% for 40 days at 75 million cells/mL ^S20. Boehringer Ingelheim reports moderate to severe retention across commercial fibres from several hundred kDa up to 0.65 µm, and far less with a larger-pore fibre ^S29. AstraZeneca links part of the problem to fibre length: in long fibres, liquid filters back in along the fibre (Starling flow), and shorter fibres or split permeate raised sieving ^{S31 S32 S35}. The Boehringer Ingelheim, WuXi polysulfone and AstraZeneca results in this section come from abstracts only ^{S28 S29 S30 S31 S32 S35}.

Fouling also ends runs. Merck & Co. found that simethicone antifoam builds up in perfusion and clogs the fibres, and that swapping filters bought only a couple of extra days ^S21. Reported fluxes are low, about 1 to 4.4 litres per square metre per hour ^{S20 S21 S24}.

Devices without a membrane

Settlers, acoustic separators, hydrocyclones and spiral channels separate cells by gravity, sound or flow instead of a membrane. The product has nothing to stick to, so these devices report little or no product retention ^{S19 S33 S38}. The trade-off is that they let some cells escape.

• Acoustic settler. Kept 86.4 to 98.9% of viable cells after infection, against 100% for ATF in the same study, and one of the two lowest values came with the highest recirculation rate; it kept dead cells less well, 81 to 98.6% ^S18. The cells were avian, not CHO, and a co-author works for the settler’s maker.
• Inclined settler. Kept 96 to 99% of cells during growth and still more than 85% late in the run as cells shrank ^S19, again with avian cells.
• Hydrocyclone. Kept 89 to 95% of mouse myeloma cells, more at higher pressure ^S23, and up to 96% of CHO cells at 50 L for 20 to 25 days at up to 50 million cells/mL ^S33. Pfizer ran one on purpose with incomplete retention in an N-1 step ^S34. All three are abstracts.
• Spiral microfluidic channel. Kept 99% of CHO cells at 4.8 million cells/mL, 91% at 22 million and 84% at 44 million with a redesigned channel ^S25: the clearest evidence that these devices lose efficiency as the culture gets denser.

In the simulator, a leaky device shows up directly: every cell that slips out through the harvest is one fewer the bleed has to remove, and below a certain retention the culture cannot hold its density at all.

Perfusion without a perfusion plant: N-1

Most companies still produce in fed-batch. Bristol Myers Squibb authors put perfusion at under 10% of commercial mammalian processes ^S09, while an industry consortium, citing earlier work, counts more than 17 launched biologics that used it over 25 years, at up to 4,000 L ^S02. The common entry point is to run only the last seed step, N-1, in perfusion, and seed the production fed-batch at a much higher density.

The results are consistent across two companies. BMS seeded fed-batch at 10 to 20 million cells/mL from an N-1 perfusion and raised titres by 100% in 10 to 14 days for four antibodies, then scaled the process to 500 L for three of them ^S09. Boehringer Ingelheim reached up to 45 million cells/mL in the N-1 step and, by seeding at 10 million cells/mL, raised titre 1.5 times in the same run time, and 1.9 times with an added lactate feed ^S10. BMS runs its N-1 at 40 pL/cell/day for 5 to 7 days and tracks densities of up to 130 million cells/mL with a capacitance probe ^S08. The attraction, as BMS puts it, is less medium than full perfusion and only small changes to existing fed-batch plants ^S09.

What perfusion buys, and what the numbers do not settle

Head-to-head numbers exist but need care. In the KTH work, perfusion harvested about 12 g of antibody against 2 g for fed-batch in the same bag over 12 days, six times more, and the authors calculate that a month at 100 million cells/mL would yield 57 g from the same 4 L working volume ^S04. The authors themselves warn that such comparisons depend on productivity, how well the fed-batch was optimised, run length and scale ^S04. Recent volumetric productivities range from 0.4 to 1.2 g/L/day at lab scale ^S17 to 0.92 g/L/day of purified antibody at 30 L pilot scale ^S05; the consortium quotes 2 to 4 g/L/day at 70 million cells/mL from earlier industry work ^S02. Whether that lowers the cost of goods depends on the case, and a 2025 cost review calls the facility fit still debated ^S15.

On the regulatory side, ICH Q13 lets a continuous batch be defined by run time, including as a range ^S01. It also notes that continuous bioreactors run longer than batch ones, so the limit on cell age has to be established for the continuous mode itself ^S01.

What it means for a plant

• Price the medium first. CSPR multiplied by target density gives the daily medium volume. A process at 50 pL/cell/day and one at 20 are different businesses.
• Count the bleed as a loss line. If a fifth or more of your product leaves with the bleed, reducing growth at steady state is worth as much as raising productivity.
• Test sieving over the full run, not the first week. Check membrane structure, pump type and fibre length before scale-up, and watch antifoam.
• If you try a device without a membrane, measure retention at your target density. Published efficiencies fall as density and flow rise, and CHO data are thin.
• N-1 perfusion is the low-risk way in. Two companies report 1.5 to 2 times the titre with existing fed-batch equipment.

Questions

How does a perfusion bioreactor work?

Fresh medium flows in continuously, a retention device lets cell-free liquid leave as the harvest, and a separate bleed removes whole culture, cells included ^S11. Once the culture reaches its target density, the operator holds it there by adjusting the bleed ^S11. In a 30 L pilot run, 1.5 vessel volumes of medium went in each day, of which 0.1 to 0.3 left with the bleed and 1.2 to 1.4 with the harvest ^S05.

Perfusion vs fed-batch: what is the difference?

A fed-batch culture adds concentrated feed and is harvested all at once after about two weeks; a perfusion culture is fed and harvested continuously for weeks. In the KTH work, perfusion harvested about 12 g of antibody against 2 g for fed-batch in the same bag over 12 days, but the authors warn that such comparisons depend on productivity, fed-batch optimisation, run length and scale ^S04. Bristol Myers Squibb authors put perfusion at under 10% of commercial mammalian processes ^S09.

What is CSPR, and what is a good perfusion rate?

The cell-specific perfusion rate (CSPR) is the fresh medium each cell gets per day, usually in picolitres. Multiply it by the cell density to get the medium flow: 15 pL/cell/day at 100 million cells/mL is 1.5 vessel volumes a day, the figure used at 30 L pilot scale ^S05. Published steady runs used between about 10 ^S17 and 70 pL/cell/day ^S03, and in two studies getting below about 30 took reinforced or purpose-built media ^{S05 S17}. Perfusion media are generally designed per process, so the right value depends on the cell line and the medium ^S17.

ATF vs TFF: which holds back more product?

Both pass culture along hollow fibres: ATF alternates the flow with a diaphragm pump, TFF drives it one way round a loop. In the KTH runs the fibres held back more antibody with TFF than with ATF ^S04. Boehringer Ingelheim traced the gap to cells broken by the peristaltic pump, and a low-shear centrifugal pump gave ATF-like results; we read that study as an abstract only ^S28. At Merck & Co., a 0.2 µm membrane with wide surface pores stayed at or above 85% sieving for 40 days while a 0.65 µm membrane with narrow surface pores fell to 70% sieving or less, so the structure of the membrane mattered more than its nominal pore size ^S20.

What is N-1 perfusion?

It runs only the last seed step before the production bioreactor in perfusion, so the fed-batch starts at a much higher density. Bristol Myers Squibb seeded at 10 to 20 million cells/mL and raised titres by 100% in 10 to 14 days for four antibodies ^S09; Boehringer Ingelheim raised titre 1.5 times in the same run time, and 1.9 times with an added lactate feed ^S10. BMS notes that it needs less medium than full perfusion and only small changes to an existing fed-batch plant ^S09.

What we could not verify

• No open-access paper reports every steady-state variable together (density, flows, growth, productivity and sieving). That is why the simulator is labelled illustrative rather than validated.
• We found no primary figure for the price of perfusion medium or its share of the cost of goods.
• Key studies on bleed recycling and on sieving mechanisms (Karst and colleagues, 2016 to 2017; Romann and colleagues, 2023) and the main techno-economic comparisons were behind paywalls.
• We found no CHO data for acoustic settlers, no centrifuge data and no primary numbers for reverse-flow TFF or TFDF. Those belong to later pieces in this series.