In Vitro Permeability Models: Caco-2, PAMPA and Transport Assays

A permeability assay returns a velocity. Apparent permeability carries units of centimeters per second and describes how fast a compound crosses a barrier the laboratory built — a monolayer of a colon carcinoma line on a plastic filter, or a lipid-soaked filter with no cells in it. Everything that makes the number useful, and everything that makes it fragile, follows from the barrier being constructed rather than borrowed.

This guide covers what Caco-2, PAMPA and bidirectional transport assays each represent, the arithmetic that turns a receiver-well concentration into a permeability coefficient, the integrity controls that decide whether a plate is reportable, what efflux ratios can resolve, and where extrapolation runs out.

What each model is a model of

Caco-2 is a human colon adenocarcinoma line that, cultured on a permeable support, differentiates into a polarized monolayer with tight junctions and an apical brush border. The reference protocol requires 21 days post-seeding to establish a stable and confluent monolayer, which is then used in a single permeability assay during the period of monolayer stability, up to day 30 ^[1]. It is colonic epithelium coaxed into small-intestinal behavior — the system's largest approximation.

PAMPA contains no cells. In the formulation examined by one group, the permeation barrier is a filter-supported dodecane–egg lecithin mixture whose molecular structure, they noted, had never been characterized. Their NMR work found that at the 9% (w/v) n-dodecane concentration used in PAMPA a purely isotropic phase is observed, corresponding to lipid aggregates 4–7 nm in diameter, and their distribution studies indicated that these reverse micelles facilitate binding to and in turn permeation across the dodecane barrier, in particular for amphiphilic solutes ^[2]. A PAMPA well is not a bilayer with defined leaflet order. It reports partition-and-diffuse behavior through a lipid-filled pore, with no transporters, no metabolism and no paracellular route.

Bidirectional transport assays are not a third model but a use of a cell model: one monolayer dosed apically, another basolaterally, the two permeabilities compared.

The apparent permeability arithmetic

The standard expression is a flux divided by the driving concentration and the area it crosses: Papp = (dQ/dt) / (A × C₀), where dQ/dt is the rate of appearance of compound in the receiver compartment, A is the filter area, and C₀ is the donor concentration at the start.

Work it through on a 12-well insert with a nominal area of 1.12 cm², 0.5 mL apical donor and 1.5 mL basolateral receiver, dosed at 10 µM and sampled to 120 min. Suppose the receiver reaches 0.24 µM and the accumulation plot is linear. Since 1 µM equals 1 nmol/mL, the receiver holds 0.24 × 1.5 = 0.36 nmol. Over 7,200 s that is dQ/dt = 5.00 × 10⁻⁵ nmol/s. The donor concentration expressed per cubic centimeter is 10 nmol/cm³. So Papp = 5.00 × 10⁻⁵ / (1.12 × 10) = 4.46 × 10⁻⁶ cm/s. The moles cancel; length over time survives.

Three traps are worth naming. The volume in the numerator is the receiver volume, because that is where the amount was measured. The concentration in the denominator is the donor concentration — not the receiver's, and not the average of the two. And C₀ must be the concentration actually delivered, assayed rather than assumed: if the compound partly precipitated, the dissolved concentration driving the flux is below nominal, so dividing by nominal C₀ divides by too large a number and the reported permeability comes out too low.

Sink conditions, depletion and mass balance

The linear formula assumes the receiver stays near enough to zero that back-flux is negligible. A common house rule holds the receiver below 10% of the donor concentration; in the example it reaches 2.4%, so the assumption holds.

Donor depletion is the quieter problem. The donor held 10 nmol/mL × 0.5 mL = 5.00 nmol, and 0.36 nmol crossed, so 7.2% of it left. If nothing else were lost the donor would sit at 9.28 µM by 120 min, a time-averaged driving concentration near 9.64 µM. Recomputing against that mean rather than C₀ raises the answer by 10/9.64 = 1.037, to 4.63 × 10⁻⁶ cm/s — small, but systematic and always the same way: using C₀ for a depleting donor understates permeability.

Mass balance catches the losses depletion arithmetic cannot see. If the end-of-experiment assay instead finds 4.30 nmol in the donor and 0.36 nmol in the receiver, recovery is (4.30 + 0.36)/5.00 = 93.2%, leaving 6.8% unaccounted for — adsorbed to plastic and filter, retained in the cells, or metabolized. Laboratories conventionally withhold rather than report a permeability value when recovery falls below about 80%, because the equation has no term for missing material.

Poorly soluble, lipophilic compounds fail this check routinely. One high-throughput Caco-2 implementation added bovine serum albumin to the buffer, selecting a lower level than the commonly used 4% in an effort to minimize the potential risk of missing the identification of efflux substrates and to avoid the extensive sample cleanup 4% BSA requires; systematic evaluation then found 0.5% BSA effective at enhancing compound solubility and reducing nonspecific binding ^[3]. That is a trade, not a fix: protein improves recovery and simultaneously changes the free concentration the barrier sees.

Monolayer integrity controls

Transepithelial electrical resistance is the routine, non-destructive integrity readout: a widely accepted quantitative technique for measuring the integrity of tight junction dynamics in cell culture models of epithelial and endothelial monolayers, performed in real time without cell damage, generally based on ohmic resistance or on impedance across a wide spectrum of frequencies, with variations arising from factors such as temperature, medium formulation and passage number ^[4].

The arithmetic is easy to get backwards. TEER = (R(sample) − R(blank)) × A: subtract the cell-free filter and medium resistance first, then multiply by area. An insert reading 620 Ω against a 120 Ω blank on a 1.12 cm² filter gives (620 − 120) × 1.12 = 560 Ω·cm². Multiplying by area rather than dividing is correct because a larger sheet of the same tissue conducts more current and so reads a lower raw resistance.

TEER and permeability are not interchangeable. TEER responds to ionic conductance through the junctions; a tracer permeability responds to one specific molecule. A study running over 150 permeability assays on Caco-2 monolayers, tracking TEER and the paracellular marker Lucifer yellow alongside nuclei and ZO-1 staining, found that manipulation of the monolayer during the assay may itself contribute significantly to Lucifer yellow flux, producing marker Papp values that depend on the sampling duration; the assay also caused a small TEER drop that fully recovered after two days back in culture medium, which let the same inserts be used on days 22, 25 and 28 ^[1]. A marker value is evidence about a particular plate on a particular day, not a property of the cell line.

Efflux ratios and what they can resolve

The efflux ratio is Papp(B→A) divided by Papp(A→B). Basolateral-to-apical is the numerator; apical-to-basolateral is the denominator. With Papp(A→B) = 0.80 × 10⁻⁶ cm/s and Papp(B→A) = 12.4 × 10⁻⁶ cm/s, the ratio is 15.5.

That single figure hides two independent measurements, each with its own error. If both directions carry a 15% coefficient of variation, the ratio inherits roughly √(0.15² + 0.15²) = 21%, so the estimate is about 15.5 ± 3.3, or roughly 12 to 19. Screening thresholds such as "efflux ratio above 2" are decisions about a noisy quantity; a ratio near the cutoff should be repeated, not rounded.

Controls define the assay more than the ratio does. One validation study characterized a Caco-2 model by determining bidirectional apparent permeabilities and efflux ratios for known P-glycoprotein substrates and a non-substrate, and considered digoxin and propranolol appropriate positive and negative controls for P-gp-mediated transport, verapamil and probenecid the corresponding controls for inhibition; reproducibility across replicate monolayers and cell passages indicated that functional P-gp expression was consistent across the passage range 25–40 they used ^[5].

Chemical inhibition is the weakest link. Zinc-finger-nuclease knockouts of MDR1, MRP2 and BCRP in a Caco-2 subclone each showed a loss of asymmetric transport for the matching probe substrates, and compared with those knockout lines standard inhibitors showed substrate-specific variation in reducing efflux ratios ^[6]. A ratio that falls when an inhibitor is added shows something moved; it does not name the transporter.

The ratio also resists conversion into an intrinsic transport rate. Investigators studying the pH dependence of efflux ratios using MDCK-MDR1 assays reported that the combined influences of paracellular transport, a previously unaccounted-for basolateral influx transporter and potential pH effects on the transporter impede the extraction of intrinsic values for active transport from the efflux ratio; they also determined that using an inhibitor affects the measurement of paracellular transport, and wrote that clear indications of transporter species preference remain elusive ^[7].

The unstirred water layer sets a ceiling

Adjacent to the monolayer sits poorly mixed buffer the compound must cross first. Treated as resistances in series, 1/Papp = 1/P(UWL) + 1/P(membrane). Take P(UWL) = 20 × 10⁻⁶ cm/s. A membrane at 200 × 10⁻⁶ cm/s gives 1/(0.05 + 0.005) = 18.2 × 10⁻⁶ cm/s, 91% of the water-layer value. Raise the membrane tenfold, to 2,000 × 10⁻⁶ cm/s, and the measured value reaches only 19.8 × 10⁻⁶ cm/s — an 8.9% change for a tenfold change in the thing being measured. Above the ceiling, the assay reports the buffer.

This is measurable, not theoretical. In a gas-lift diffusion cell holding Caco-2 monolayers, In a gas-lift diffusion cell holding Caco-2 monolayers, testosterone permeability was 4.07, 10.90 and 14.18 × 10⁻⁵ cm/s at mixing flow rates of 0, 15 and 40 mL/min — a 3.48-fold rise from the unstirred condition to the fastest — with mannitol flux staying at least 15-fold below testosterone flux, which the authors read as indicating the stirring flow rates used had not affected monolayer integrity; the apparent water layers were calculated to be 1,966, 733 and 564 µm at those three flow rates, and a stagnant Transwell gave 3.08 × 10⁻⁵ cm/s with a calculated layer of 2,597 µm ^[8]. Those thicknesses were derived from the permeabilities themselves, so the inverse relation between permeability and calculated thickness follows from the calculation rather than testing it independently. The load-bearing result is the direct one: stirring alone, with the barrier unchanged, raised the measured permeability 3.48-fold.

The layer can also hide transport. Working with grepafloxacin, whose permeability coefficient is high, investigators found that transporter inhibitors did not alter secretory transport in conventional Transwell Caco-2 experiments; after agitating the solutions to thin the water layer and lowering the temperature, the layer was no longer rate-limiting and the inhibitor effects showed up as decreased secretory and increased absorptive apparent permeability. They concluded that careful analysis is required to evaluate transporter contributions whenever the water layer affects apparent permeability ^[9].

Nor is the layer symmetric. Using a purpose-built stirring apparatus, one group found the unstirred layer thicker in a Transwell than in a side-by-side chamber even under vigorous basolateral magnetic stirring, calculated the apical layer to be thicker than the basolateral one, and obtained different cellular permeability coefficients from the two apparatuses; they concluded that an orbital shaker does not produce symmetric hydrodynamics in both chambers and that the asymmetric layer may complicate exact analysis of polarized transport ^[10]. An efflux ratio built from two directions with unequal water-layer resistances is not a clean ratio.

PAMPA arithmetic is not the same arithmetic

PAMPA is rarely analyzed with the linear initial-rate formula. Both compartments are finite and approach equilibrium during the incubation, and a substantial fraction of compound can be retained in the lipid barrier, so implementations use equations accounting for the approach to equilibrium and, usually, for membrane retention. Laboratories choose differently and the results are not interchangeable: a meta-analysis of public permeability databases noted that for PAMPA two different permeabilities are reported, intrinsic and apparent, that these can vary in degrees of magnitude, and suggested being extra cautious in using literature permeability data ^[11]. Barrier composition varies as well; the same meta-analysis describes PAMPA as a cell-free model system formulated to mimic different barriers, gastrointestinal epithelia or skin among them ^[11].

Stirring is a design variable in PAMPA too. In one comparison of 17 fluoroquinolones, including three congeneric series with systematically varied alkyl chain length, the PAMPA method employed stirring adjusted so the unstirred water layer thickness matched the 30–100 µm range estimated to be in the human small intestine ^[12].

Peptides break several assumptions at once

A peptide crossing a Caco-2 monolayer can be hydrolyzed on the way. In a study of five milk protein-derived peptides and a whey protein isolate digest, only a small percentage of what was added apically crossed intact — 0.05% for LPYPY up to 0.47% for WR — and all five were susceptible to peptidase action during the transport study ^[13]. Two consequences follow for the arithmetic. The analytical method must quantify the intact sequence specifically, or the numerator counts fragments; and mass balance reads low unless fragments are measured too, so poor recovery may reflect degradation rather than adsorption.

The limits of extrapolation

Correlations between assays, and between assays and in situ data, are moderate and depend on what has been corrected. Across that fluoroquinolone set the rat–PAMPA correlation was r² = 0.87, rat–Caco-2 r² = 0.63 and Caco-2–PAMPA r² = 0.66, the last improving to r² = 0.82 once the Caco-2 data were corrected for the unstirred water layer ^[12]. Correcting one axis for the water layer raised the Caco-2–PAMPA agreement by 0.16 in r², a shift smaller than but comparable to the 0.24 separating the rat–Caco-2 and rat–PAMPA correlations. How the data are corrected moves agreement on a scale not far from which model produced them.

The meta-analysis of published data is blunter. Evaluating four experimental and two computational methods, it reported that repeatability of reported permeability is not great even for the same method; that the best agreement was between cell-based methods, and between black lipid membrane data and computational predictions; that unstirred water layer permeability limits agreement between water-layer-limited methods and those that are not; and that cell-based methods provide results only in a small range, roughly 10⁻⁸ to 10⁻⁴ cm/s. Its authors recommended caution in interpreting and combining permeability data, in-house or published ^[11].

What is not established matters as much. There is no general conversion factor turning a Caco-2 or PAMPA number into a quantity in an organism, and none of these assays measures anything that happens in a person. Nor is there a settled threshold above which a monolayer counts as tight enough: acceptance criteria for TEER and marker permeability are laboratory conventions set against historical control data on that clone, those inserts and that medium, not values derived from any primary study. The same holds for the 80% recovery rule and the 10% sink criterion used above. They make numbers comparable within a laboratory, not between laboratories.

So a permeability coefficient is interpretable only alongside its conditions: clone and passage, days post-seeding, insert type and membrane area, buffer and pH on each side, protein content, shaking speed, incubation and sampling times, the donor concentration as assayed, TEER and marker permeability before and after, recovery, and the direction of every value in the table. Absent that, two numbers sharing the same units are not describing the same experiment.