ATF vs TFF in perfusion: what the filter holds back, why it fouls and how it scales

Research 46 primary sources read, 41 cited: papers, one doctoral thesis, an industry design basis and vendor documents 2 Oct 2026

73% and 71%

A perfusion bioreactor needs a device that keeps the cells in and lets the spent medium out, with the product in it. In the antibody processes covered here, that device is a cartridge of hollow fibres, run in one of two ways: alternating tangential flow (ATF) or tangential flow filtration (TFF). Which one fouls less, and which one holds back less product, is argued by vendors and users alike. We read 46 primary sources, from 2011 to 2026, to see what has actually been measured. This is the third piece of our perfusion series: how a perfusion culture works is in Perfusion, explained, and how it compares with fed-batch is in Fed-batch vs perfusion.

The seven findings in short

• ATF and TFF use the same filter. What differs is the pump. An ATF moves the culture back and forth with a diaphragm pump through one connection. A TFF circulates it one way around a loop. Confidence: high.
• Both hold back product, and more as the filter ages. In two studies antibody transmission fell to 50% or less within 10 to 23 days, and an industry consortium writes that it can typically reach 50 to 60% after 20 days. Confidence: high.
• Whether a well-built TFF matches an ATF is not settled. With a peristaltic pump the TFF did worse in two studies. With a low-shear pump one study found them equal, 73% and 71% recovery, and another found the TFF retaining up to 50% of the antibody against about 10%. Confidence: medium at best. Three laboratories, bench scale.
• Product held back is delayed, not lost, unless cells are bled. Without a bleed, 86 to 92% of the antibody was harvested although transmission had fallen to 50% or less. With a daily bleed it was 54 to 71% with ATF and 46 to 56% with TFF. Confidence: medium, one study, but the mechanism is a mass balance.
• What fouls the fibre is particles about the size of the pores, not whole cells. Roughly 20 to 200 nm. Two sources tie them to cells breaking, a third points to antifoam, and what they are made of is disputed. Confidence: high for the size, medium for the link with cell lysis, low for the identity.
• Membrane structure matters more than the pore rating, and more crossflow is not a cure. One 0.2 µm membrane with wide surface pores kept transmission above 85% for 40 days. A larger rating, 0.65 µm, did worse in three of the four comparisons we read. Confidence: medium; the membrane structure result comes from one company.
• Scale and cost are the thin part. Past 500 to 1,000 L an ATF setup needs several units, and a TFF needs longer fibres that retain more. No source gives a filter price or a measured filter life at production scale. Confidence: low.

How we built this

We read 25 sources in full and 21 from the abstract only, and we say so where it matters. The full-text set includes papers, one doctoral thesis and six documents published by vendors; several of the papers also have vendor co-authors. Who is talking matters here. Repligen makes the ATF and the tangential flow depth filter and is the author of five documents ^{S20 S21 S22 S23 S24} and a co-author of two papers ^{S16 S17}. Levitronix makes the low-shear pump used in most recent TFF work; it co-authored a pump report ^S25 and two papers we read as abstracts ^{S38 S39}, and the copy of the Boehringer Ingelheim paper we read ^S04 is hosted on its website, although the authors are all from Boehringer Ingelheim. Cytiva, Artemis Biosystems and MilliporeSigma have co-authors in ^S09 and ^S43, in ^S14 and in ^S40. The 2013 KTH study was financed by GE Healthcare, which has co-authors on it and makes the bioreactor and the fibre that were used ^{S01 S02}. Two papers are an industry consortium’s design assumptions, not measurements ^{S12 S13}. For the ETH Zurich comparison we read the author’s doctoral thesis ^S03, not the journal article, which was behind a paywall. The three head-to-head comparisons of ATF and TFF that we read in full were run at 1.5 to 4 L, with the two systems side by side for 30 days or less; a fourth, at 200 mL, we read only as an abstract ^S42.

What tangential flow filtration is

In ordinary filtration the liquid is pushed straight at the filter, and whatever does not pass stays on it. In tangential flow filtration the cell suspension flows along the membrane, which may prevent fouling ^S01: the flow creates shear that clears the layer forming on the surface ^S04. The same paper states the price. More shear cleans better, but it can also harm the cells ^S04.

In perfusion the membrane is a bundle of hollow fibres inside a cartridge. The culture flows inside the fibres, in the lumen ^S01, and cell-free liquid, the permeate, leaves through the fibre wall. The 2013 KTH study, the earliest side-by-side comparison we read, used 50 polysulfone fibres with 0.2 µm pores and a 1 mm lumen, 60 cm long, with 850 cm² of area ^S01. An ATF2 cartridge described in a modelling paper had 75 fibres of 20 cm and 470 cm² ^S05. The vendor’s range for the ATF is polyethersulfone or polysulfone, at 0.2 µm, 0.5 µm or 50 kD, with a 1 mm inner diameter ^S21.

The flow through the wall is very small. It is quoted as about 2 litres per square metre per hour (LMH) “traditionally” ^S05, as 1 to 2 LMH in a Merck & Co. study ^S06 and, in a paper that cites earlier work for it, as typically below 3 LMH for ATF, which is why “large membrane areas are required” ^S08. Higher values appear in an intensified Merck & Co. process, about 4.4 LMH ^S07, and in a vendor test, 5.7 LMH ^S22. The crossflow along the fibre is set as a shear rate: 1,000 to 3,400 per second at KTH ^S02, below 2,000 at Boehringer Ingelheim ^S04, 1,250 to 3,000 at Merck & Co. ^S06.

One detail matters later. The pore rating on the label is not what the culture meets. Under the electron microscope, a membrane rated 0.2 µm had surface pores of 0.5 to 2 µm, and one rated 0.65 µm from the same maker had surface pores of 0.3 to 0.6 µm ^S06.

Same filter, two pumps

Much of the setup, hollow fibre included, is the same in both systems. The difference is the pump that drives the flow along the fibres ^S04.

In an ATF, a diaphragm pump moves the culture in and out of the fibres through a single connection to the bioreactor ^S03. Air pressure pushes the diaphragm and sends liquid through the fibres back to the bioreactor. Vacuum pulls the diaphragm the other way and draws liquid from the bioreactor into the fibres. Two strokes make one cycle ^S21. The vacuum is not optional: the vendor’s guide says that pressure from the bioreactor is not enough to deflate the diaphragm completely ^S21. With a 100 mL pump, one cycle took 11 seconds at 1.0 L/min and 8 seconds at 1.5 L/min ^S03.

In a TFF, a pump circulates the culture one way around a loop, which takes two ports on the bioreactor instead of one ^S04. The loop pump was peristaltic in the early work ^{S01 S04} and is a centrifugal pump made by Levitronix in later studies ^{S03 S06}, described as magnetically levitated ^{S03 S04}. In the fibre, the velocity is constant in a TFF and swings between a maximum and zero in an ATF ^S11. The cells spend a short time outside the bioreactor: under a minute in the 250 mL loop of the 2013 study ^S01, 8 seconds in the control runs of a Merck & Co. study ^S07, and 2 to 7 seconds in an industry design for 500 to 2,000 L, which notes that this is near the time to hypoxia ^S12.

The claimed advantage of the ATF is the backflush. Each change of direction is said to push some permeate back through the pores and clean them ^{S03 S31}. The evidence for it is thin. The 2013 KTH paper noted that the effect “has not been quantified” ^S02. A computational model of one ATF fibre found that part of the permeate flows back into the fibre, which is known as Starling flow, with the direction of the flow through the wall changing about halfway along. It concluded that at 2 LMH a one-way flow would use half of the membrane while the alternating flow uses all of it over a cycle. In the same paper, Starling flow did not reduce fouling as the crossflow increased ^S05. For one-way TFF, a later model that we read as an abstract treats Starling flow as a major contributor to fouling and product retention, not as a cure ^S30.

Both hold back product, and more with time

The filter is there to stop cells. It also stops part of the product. Papers report this as sieving, the product concentration in the permeate divided by the concentration in the bioreactor ^S04, or as retention, its complement.

At KTH, with a 0.2 µm fibre and the culture held between 20 and 35 million cells/mL by daily bleeds ^S01, the ratio was 100% during the first days and then decreased rapidly ^S02. The harvest settled at 93 to 158 mg/L while the bioreactor held 176 to 243 mg/L ^S02. In a long TFF run of the same study, after day 20, when the cell density was raised to about 100 million cells/mL ^S01, the harvest was about 154 mg/L while the bioreactor reached 1,000 mg/L or more; the first cartridge of that run was in use for 30 days ^S02. WuXi Biologics, with a bispecific antibody on its own platform, saw retention rise gradually from 20% to 60% in a run harvested from day 7 to day 23 ^S10. An industry consortium writes that transmission “typically can reach as low as” 50 to 60% after only 20 days; that is a description of practice, not a measurement of its own ^S13. Boehringer Ingelheim tested commercial fibres from several hundred kilodaltons to 0.65 µm and reports moderate to severe retention with all of them (abstract) ^S26.

Three of the studies we read in full ran ATF and TFF side by side on the same fibre. They do not agree.

KTH, 2013. The TFF used a peristaltic pump. Antibody was partially retained by both systems “but more severely using TFF”. In the runs held at 25 million cells/mL, the share of all antibody produced that reached the harvest was 54 to 71% with ATF and 46 to 56% with TFF ^S02.

ETH Zurich, 2016. The TFF used a magnetically levitated centrifugal pump, the fibre had 0.5 µm pores and both loops ran at 1.5 L/min. The ATF retained 10% of the antibody on average (two runs). The TFF, in a single run, retained up to 50% when the cell density was raised from 20 to 60 million cells/mL; the conclusion of the same thesis chapter says up to 60%. The author’s explanation is the backflush of the ATF, which was not measured ^S03.

Boehringer Ingelheim, 2017. Two experiments on a 0.2 µm fibre. With a peristaltic pump, the TFF membrane blocked completely on day 15, while the ATF kept sieving above 50% for the 20 days of culture. With a low-shear centrifugal pump, the TFF ran the full 30 days and total recovery was 73% in the TFF against 71% in the ATF. The authors conclude that the ATF’s advantage came from its diaphragm pump, which puts less mechanical stress on the culture, and not from the back and forth flow. They add that sieving was “alleviated but not solved” in either system, and that their result is “in stark contrast” to the ETH one ^S04.

We cannot say which of the last two is right. The fibres differed (0.5 µm against 0.2 µm), the ETH study ran at a very low flux of 0.4 LMH ^S03, the cell lines were different, the ETH result rests on one TFF run, and we found no repeat of either study.

Held back is not lost, unless there is a bleed

What retention costs depends on what happens to the culture that stays behind. With no bleed, product that does not pass the filter today stays in the bioreactor, its concentration rises, and more of it passes later. In a KTH run without bleeds, the ratio of harvest to bioreactor concentration fell to 50% or less after 10 days, yet 86 to 92% of all the antibody made was harvested ^S02. A vendor poster makes the same point from its own data: even at 60% retention, more than 95% of the protein was recovered in the harvest ^S22. The poster does not say whether that culture was bled.

With a bleed it is different. Culture is removed every day to hold the cell density, and the product dissolved in it goes too. In the words of the KTH authors, the main loss caused by retention “was product discarded from the bioreactor in the daily bleeds” ^S02. That is where the 54 to 71% and 46 to 56% above come from. The authors count a second, smaller loss: the antibody still in the bioreactor when the run ends ^S02. The ETH thesis says the same: the yield in the harvest is directly linked to the loss of product in the bleed ^S03.

A steady-state mass balance shows the size of the effect. If s is the sieving, H the harvest flow and B the bleed flow, the share of product that reaches the harvest is s × H / (s × H + B). With a bleed of 20% of the outflow, that is 80% when the filter passes everything, 67% when sieving is 50% and 44% when it is 20%. With no bleed it tends to 100% whatever the sieving. This is our calculation, for a well-mixed bioreactor with constant sieving, and not a validated model: the KTH figures are consistent with it, but they are ranges over several runs and the bleed volumes were not reported, so they cannot confirm it. For scale, a pilot run at 100 million cells/mL bled 0.1 to 0.3 vessel volumes a day, 6.7 to 20% of the perfusion flow ^S09.

Two more consequences. Retained product stays longer in the bioreactor, which the KTH authors flag as an issue for sensitive glycoproteins ^S02. And the consortium’s design for a continuous plant simply assumes a 90% yield at the cell retention step ^{S12 S13}.

What fouls the fibre

Not the cells. In the Boehringer Ingelheim study most particles in the permeate were below 10 nm, while particles of about 100 nm accumulated in the bioreactor as the culture went on. The team then split a culture three ways and ran each fraction over a new 0.2 µm fibre. Permeate alone, and cells resuspended in permeate, caused no significant loss of transmission. The supernatant, with its 100 nm particles, caused a loss that was “severe and almost instantaneous”. Lysed cells produced far more of these particles than apoptosis alone ^S04. A later abstract from the same company gives about 20 to 200 nm ^S26, a range that a Merck & Co. paper takes up ^S06. The vendor’s poster gives 70 to 200 nm and also links retention to cell lysis ^S22.

What the particles are made of is open. The candidates, and who puts them forward:

• Extracellular vesicles. At Merck & Co., the amount of vesicles of 50 to 200 nm in the permeate fell before the transmembrane pressure rose and sieving dropped. On used fibres, the foulant patches were rich in cellular material but not in antifoam (abstract) ^S35.
• Antifoam. A later Merck & Co. paper found that the antifoam does build up in the bioreactor, more than 10 fold between day 6 and day 24, to about 250 to 500 ppm. In a small-scale filter test, 75 to 250 ppm raised the pressure and cut sieving, and 500 ppm blocked the filter at once. A culture given twice the antifoam lasted 14 days ^S07. The vendor’s poster reports no effect of antifoam alone in a 30 day cell-free test ^S22.
• Host cell protein. On fibres from industrial ATF runs, the deposit was mostly host cell proteins with some antibody (abstract) ^S40.
• DNA. Named the main soluble foulant in a scale-down study (abstract) ^S43.

The authors say it themselves: the factors causing filter fouling “remain largely unknown” ^S07, and the mechanism “remains poorly understood” ^S13. The 2025 Merck & Co. paper adds that the causes of fouling may differ from those of sieving ^S07.

In one study, the pump did the damage

In the first Boehringer Ingelheim experiment, the death rate measured by LDH release was 0.22 per day in the TFF with a peristaltic pump and 0.09 per day in the ATF. The cells in the TFF appeared to grow more slowly, 0.32 against 0.49 per day, and the authors put that down solely to the higher rate of lysis. Viability by trypan blue showed a smaller gap, 81.1% against 91.5% on day 14. By day 8 the TFF bioreactor held about 4.5 times more particles than the ATF one ^S04.

A batch culture looped through a peristaltic pump pointed to the cause: its death rate was 0.13 per day against 0.02 without the loop, and it peaked at 4.7 million cells/mL against 12.2 million. With a centrifugal pump in the TFF, death rates and particle counts were the same in both systems ^S04.

Not every study agrees on the pump. KTH saw no cell damage from its peristaltic pump at up to 2.75 L/min with the fibre bypassed, in tests of up to three days ^S02. At ETH Zurich, a stress probe made of particle aggregates, run in water without cells, showed a much lower maximum stress in the ATF than in the TFF with a centrifugal pump at comparable flow; the stress was dominated by the pump, and the TFF had to be limited in flow to stay under the 32 Pa that this cell line tolerated ^S03. A pump report co-authored by Levitronix, not peer reviewed, gives death rate constants, from 12 hour tests, of 0.023 per hour for its levitated pump against 0.03 for a peristaltic pump and 0.036 for a four-piston diaphragm pump ^S25, which is not the air-driven diaphragm of an ATF ^S21.

The ATF has its own stress routes. Sanofi links estimated residence times above 75 seconds with lower growth and productivity (abstract) ^S32. Inside the fibre it is gentler on average: by a calculation in a KTH paper, the time-averaged shear in an ATF is 2/π, about 0.64, of that in a TFF with the same peak ^S11.

What helped and what did not

Membrane structure. At Merck & Co., a 0.2 µm membrane with wide surface pores kept antibody sieving above 85% throughout 40 days, whatever the shear rate and flux studied, and averaged 98% in a 29 day run at 74 million cells/mL. A 0.65 µm membrane from the same maker ended at 71% and 55%, and those cultures were stopped after three weeks ^S06. This is the pair described earlier: the membrane rated 0.2 µm was the one with the wider surface pores.

A larger pore rating is not a reliable fix. At WuXi Biologics, product retention was 38% higher with 0.65 µm fibres than with 0.2 µm ones (abstract) ^S27. At AstraZeneca, 0.2 µm filters kept higher sieving than 0.65 µm ones (abstract) ^S29. The vendor’s poster reports “slightly improved protein retention” with 0.65 µm ^S22, which we read as slightly less retention. WuXi’s reading is that it depends on the debris: larger pores may help when particles of about 0.01 to 0.2 µm dominate, and 0.2 µm pores when most particles are larger than that ^S27. Boehringer Ingelheim, after finding retention with every commercial fibre up to 0.65 µm, reports that a fibre with an “unconventionally larger” pore size reduced it drastically; the abstract does not give the size (abstract) ^S26. Going much larger changes the trade: a macroporous membrane of 1 to 4 µm delivered 13,000 L/m² of throughput, but its permeate was turbid and needed a depth filter before chromatography (abstract) ^S41.

Shorter and wider fibres. AstraZeneca found higher sieving with the shortest fibres and with larger inner diameters (abstract) ^S29. At Merck & Co., two fibres fed by the same bioreactor gave 70% and 46% transmission after 21 days; the worse one had longer, narrower lumens and ran at the higher shear rate ^S06.

More crossflow did not help. At KTH a higher recirculation rate did not improve the ratio ^S02. At Merck & Co. it led to a lower final sieving on the 0.65 µm membrane, in a run that also had twice the flux ^S06.

Backflush and flux. In a scale-down test with antifoam at Merck & Co., doubling the crossflow lowered the transmembrane pressure by about 35%, and reversing the permeate flow brought the pressure back to baseline ^S07. Sanofi reports that above 60 L/m² per day, about 2.5 LMH by our conversion, sieving fell by more than 1% a day (abstract) ^S32.

The forms of TFF

“TFF” now covers several designs. What has been measured for each is uneven.

The last row needs a caution. The vendor’s data sheet claims “negligible product retention” for TFDF and sizes it for antibodies at 3 cm² per litre of process volume, but shows no antibody data ^S23, and we found no peer-reviewed data for a long antibody perfusion run. The published work we read is on viral vectors ^{S15 S16 S17}.

Open-pore designs also change what goes downstream. With a tubular membrane of about 10 µm, protein and DNA concentrations were more or less the same in the vessel and in the permeate, while the hollow fibre held most of them back ^S14. The macroporous membrane above needed a depth filter for the same reason ^S41.

Scaling up

The vendor’s data sheet lists three production ATF sizes: ATF 4 for 10 to 50 L of working volume with 0.77 m² of filter, ATF 6 for 50 to 200 L with 2.5 m², and ATF 10 for 200 to 1,000 L with 11 m². Their nominal maximum perfusion rates are 50, 200 and 1,000 L a day ^S20. Dividing one by the other gives 2.7, 3.3 and 3.8 LMH (our calculation): the flux changes far less than the area, which grows with the volume.

In practice, the consortium reports that ATF systems are often used for 500 L bioreactors and below, that two or more may support 1,000 L, “but TFF is the preferred choice”, and that some of its authors have used as many as four on a 2,000 L bioreactor ^S13. Its design basis for 2,000 L at 1.5 vessel volumes a day calls for 30 m² of membrane, as three ATF 10 units or as TFF modules four in series. The modules go in series to limit Starling flow; in parallel, the loop would need 500 to 1,000 L/min ^S12. That design works out at 100 L/m² per day, about 4.2 LMH (our calculation), above the 60 L/m² per day at which Sanofi saw sieving decline in one ATF process (abstract) ^S32.

Each system has its own ceiling. For the ATF it is the pump: AstraZeneca writes that the few diaphragm pumps on the market prevent the use of vertical space when scaling beyond 500 L, and that stacked filters with viscous culture demand more vacuum than a facility supplies (abstract) ^S28. KTH saw the mechanism at bench scale: at 132 million cells/mL the alternating flow stopped, possibly because the vacuum could no longer pull the viscous broth ^S01. For the TFF it is the fibre: longer filters at scale worsen the retention caused by Starling flow (abstract) ^S28, and a 2026 paper with authors from several companies, a pump vendor among them, says that scaling TFF perfusion still “often requires time-consuming trial-and-error efforts” (abstract) ^S38.

On filter life there is little. At bench scale in 2013, single cartridges were in use for 30, 14, 22 and 27 days, and only the first is reported as replaced because of probable fouling; the authors say the data do not support a general conclusion ^S01. Merck & Co. measured 900 to 1,200 L/m² of throughput for its 0.2 µm membrane ^S06. The consortium assumes 1,000 L/m² and two sets of membranes per 20 day run ^S12. For an intensified process at 85 to 100 million cells/mL, Merck & Co. writes that filters can be swapped once during a run at large scale, and that the swap only extends the culture by a couple of days ^S07. The one manufacturing-scale case we found is a 3,000 L seed culture at Biogen: pressure rose on the last day, and a 30% lower permeate flow cut it by 75% (abstract) ^S37.

Cost: not resolved

No source we read gives the price of a perfusion hollow fibre, the filter cost per run or the value of the product lost to retention. What exists are the design assumptions above: 1,000 L/m², two membrane sets per run and a 90% yield at the retention step ^S12. For one 2,000 L bioreactor that is 60 m² of membrane per 20 day run (our calculation). The only price we found is a vendor’s assumption of 5,000 dollars per square metre in a cost model for lentivirus filters ^S24, which is not a quote for hollow fibres, and we do not build a figure on it. A 2013 model from University College London estimated that ATF perfusion could save 20% of the cost of goods against fed-batch if it reached five times the cell density (abstract) ^S46; it does not compare ATF with TFF. We cannot say which of the two is cheaper to run.

A hollow fiber bioreactor is something else

The term is easy to confuse with the filter. In a hollow fiber bioreactor the cells live inside the cartridge. In the two examples we read they sit outside the fibres, in what is called the extracapillary space. Medium flows through the fibre lumens, and nutrients and oxygen reach the cells through the fibre wall ^S19. The cartridge is the bioreactor, and it is small: one used for T cells had six fibres and 20 cm² ^S19; in another, used to produce a virus, about 20 mL was harvested from the extracapillary space every second day ^S18. In the perfusion systems of this article it is the other way round. The cells pass through the fibre lumens and go back to a separate bioreactor ^S01.

What it means for a plant

• Ask for sieving over time, not for the pore rating. Two membranes with the “wrong” order of ratings gave 98% and 55 to 71% ^S06. A sieving curve over the full run length, at the intended cell density, is the number to request.
• If you choose TFF, the pump and the loop are the design. A peristaltic pump cost a run on day 15 in one study; a low-shear pump on the same type of fibre matched the ATF ^S04.
• Measure lysis directly. The damage showed in LDH release and in particle counts; the viability gap by trypan blue was smaller ^S04.
• The bleed sets what retention costs. With no bleed, retained product comes out later. With a bleed, part of it is discarded every day ^S02.
• Watch the antifoam. It did not leave through the filter and built up more than 10 fold in one process ^S07.
• Plan for filter changes. The consortium design carries two membrane sets per 20 day run ^S12, and Merck & Co. writes that a swap extends an intensified culture by only a couple of days ^S07.
• Expect a different problem at scale. Several ATF units and their vacuum demand past 500 to 1,000 L ^{S13 S28}; modules in series and loop design for TFF ^S12.

Questions

What is tangential flow filtration?

A way of filtering in which the liquid flows along the membrane instead of straight at it, so that the flow itself sweeps the surface ^{S01 S04}. In perfusion cell culture the membrane is a cartridge of hollow fibres: the culture flows inside the fibres and cell-free liquid leaves through their walls at a very low flux, from about 1 to 4.4 LMH in two Merck & Co. processes ^{S06 S07}.

How does alternating tangential flow work?

A diaphragm pump moves the culture in and out of the hollow fibres through one connection to the bioreactor ^S03. Air pressure pushes the liquid back to the bioreactor and vacuum draws it into the fibres; two strokes make a cycle, which took 8 to 11 seconds in one bench setup ^{S21 S03}.

ATF vs TFF: which holds back more product?

With a peristaltic pump, the TFF held back more in two studies ^{S02 S04}. With a low-shear pump the evidence is split: equal recovery, 73% and 71%, in one study ^S04, and up to 50% retention in the TFF against about 10% in the ATF in another ^S03. Both retain more as the filter ages.

What is a sieving coefficient?

The product concentration in the permeate divided by the concentration in the bioreactor. At 100% all the product passes; at 0% none does ^S04. Retention is 100% minus sieving.

Is a larger pore size better for product transmission?

Not reliably. A 0.65 µm rating did worse than 0.2 µm in three of the four comparisons we read ^{S06 S27 S29} and was reported as slightly improved in one, the vendor’s ^S22. In one Merck & Co. study, the structure of the membrane surface explained more than the rating ^S06.

What is a hollow fiber bioreactor?

A cartridge that is itself the culture vessel: the cells grow in it, outside the fibres in the two examples we read, while medium flows through the fibres ^{S19 S18}. It is not the same as a hollow fibre filter on a perfusion bioreactor, where the cells only pass through the fibres and return to the vessel ^S01.

What we could not verify

• No head-to-head comparison of ATF and TFF beyond 4 L, none with the two systems side by side for more than 30 days, and no repeat of the two studies that disagree. One more comparison, at 200 mL, was read only as an abstract ^S42.
• No sieving curve printed as numbers: every time course is a figure, so we quote start and end values only.
• The ETH comparison was read in the author’s thesis, which gives 50% in its results and 60% in its conclusion for the same TFF run ^S03.
• Reverse-flow, co-current and stacked-filter TFF were read as abstracts; their gains are relative and the baselines are not given ^{S42 S39 S28}.
• No peer-reviewed antibody perfusion data for TFDF.
• No measured filter life for a production bioreactor of 500 L or more, and no filter prices.
• We did not build a model of sieving over time. No source prints the numbers needed to test one, so the only calculation of ours in this article is the steady-state balance, labelled as such.