Abstract
A P-glycoprotein (P-gp) IC50 working group was established with 23 participating pharmaceutical and contract research laboratories and one academic institution to assess interlaboratory variability in P-gp IC50 determinations. Each laboratory followed its in-house protocol to determine in vitro IC50 values for 16 inhibitors using four different test systems: human colon adenocarcinoma cells (Caco-2; eleven laboratories), Madin-Darby canine kidney cells transfected with MDR1 cDNA (MDCKII-MDR1; six laboratories), and Lilly Laboratories Cells—Porcine Kidney Nr. 1 cells transfected with MDR1 cDNA (LLC-PK1-MDR1; four laboratories), and membrane vesicles containing human P-glycoprotein (P-gp; five laboratories). For cell models, various equations to calculate remaining transport activity (e.g., efflux ratio, unidirectional flux, net-secretory-flux) were also evaluated. The difference in IC50 values for each of the inhibitors across all test systems and equations ranged from a minimum of 20- and 24-fold between lowest and highest IC50 values for sertraline and isradipine, to a maximum of 407- and 796-fold for telmisartan and verapamil, respectively. For telmisartan and verapamil, variability was greatly influenced by data from one laboratory in each case. Excluding these two data sets brings the range in IC50 values for telmisartan and verapamil down to 69- and 159-fold. The efflux ratio-based equation generally resulted in severalfold lower IC50 values compared with unidirectional or net-secretory-flux equations. Statistical analysis indicated that variability in IC50 values was mainly due to interlaboratory variability, rather than an implicit systematic difference between test systems. Potential reasons for variability are discussed and the simplest, most robust experimental design for P-gp IC50 determination proposed. The impact of these findings on drug-drug interaction risk assessment is discussed in the companion article (Ellens et al., 2013) and recommendations are provided.
Introduction
In recent years, the role of membrane transporters in the absorption, disposition, and excretion of drugs has been increasingly recognized. In particular, P-glycoprotein (P-gp; encoded by the MDR1 or ABCB1 gene in human) has been shown to impact drug pharmacokinetics by limiting oral absorption, restricting central nervous system penetration, and promoting excretion. For drugs that are transporter substrates and are not significantly metabolized, such as talinolol and digoxin, P-gp or other transporters play an important role in absorption and disposition. This may lead to drug-drug interactions (DDIs) when coadministered with other drugs that also interact with these transporters (Schwarz et al., 2000; Juan et al., 2007; Fenner et al., 2009; Shirasaka et al., 2010). Digoxin has a narrow therapeutic window; consequently, even slight changes in plasma exposure have been associated with adverse events. As a result, many examples of clinical digoxin DDI studies have been reported (Fenner et al., 2009) in which the mechanism of interaction is frequently ascribed to P-gp inhibition and sometimes to P-gp induction.
The recent DDI draft guidance from the FDA (US FDA/CDER, 2012; http://www.fda.gov/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/ucm064982.htm) provides decision criteria to assess the risk of a clinically significant DDI resulting from P-gp inhibition. A clinical DDI study with digoxin is recommended when the maximum total plasma (bound plus unbound) concentration of the investigational drug at steady state ([I]1) divided by its in vitro P-gp inhibitory potency (IC50) is greater than or equal to 0.1 or, for orally administered drugs, its nominal gut concentration ([I]2) divided by its IC50 is greater than or equal to 10.
These decision criteria, originally proposed by Zhang et al. (2008) and reinforced by Agarwal et al. (2013), are based on in vitro P-gp IC50 data without regard to experimental system or remaining transport activity equation and where each IC50 value is generated by one single laboratory only. In both articles, the authors emphasized the need for standardization of in vitro methods to ensure that the most appropriate decision criteria are established. Two additional articles proposed different decision criteria that are based on IC50 values for multiple compounds generated in one single laboratory, using a single experimental system and a single transport activity equation. Cook et al. (2010) using human colon adenocarcinoma cells (Caco-2) cells proposed cut-off values for [I]1/IC50 > 0.1 and for [I]2/IC50 > 5 using the net-secretory-flux equation, while Sugimoto et al. (2011) using Lilly Laboratories Cells—Porcine Kidney Nr. 1 cells transfected with MDR1 cDNA (LLC-PK1-MDR1) cells proposed a cut-off value for [I]2/IC50 > 30, using an efflux ratio-based equation.
A variety of systems are available to measure P-gp activity including inhibition of efflux of probe substrate from P-gp-expressing cells, inhibition of unidirectional or bidirectional transport across a monolayer of P-gp-expressing polarized cells, and inhibition of uptake into inside-out membrane vesicles prepared from P-gp-expressing cells. The pharmaceutical industry frequently utilizes P-gp expressing polarized cell lines such as Caco-2, Madin-Darby canine kidney cells transfected with MDR1 cDNA (MDCKII-MDR1), and LLC-PK1-MDR1 to measure P-gp mediated transport and to assess potential DDIs.
In addition to the use of various cellular and vesicular systems to evaluate inhibition of P-gp, equations used to calculate remaining P-gp transport activity as a function of inhibitor concentration also vary between laboratories and have been based on efflux ratio, net-secretory-flux, or unidirectional flux [Tang et al., 2002; Kalvass and Pollack, 2007; Balimane et al., 2008; and the FDA draft guidance for drug interactions (US FDA/CDER, 2006, 2012)]. To date, limited data are available to appropriately compare IC50 values generated using different experimental systems and transport activity equations. Therefore a collaborative effort was initiated between 23 pharmaceutical and contract research laboratories and one academic institution to compare IC50 data generated using four in vitro systems (Caco-2, MDCKII-MDR1, LLC-PK1-MDR1 cells, and MDR1-expressing membrane vesicles) and six transport activity equations. Multiple commercial nonlinear regression software packages were employed to fit the inhibition curves. The objective of this collaboration was to determine in vitro IC50 values for 16 drugs for which clinical digoxin DDI data are available, to assess the extent of variability in IC50 values obtained in different systems and using different transport activity equations and to discern whether any one system or transport activity equation outperformed the others, both in terms of robustness of the assay result (IC50 value) and assessment of DDI risk. In this article, we focus on the variability in IC50 values. The DDI risk analysis and recommendation for new decision criteria is presented in the companion article.
Materials and Methods
Chemicals.
Sixteen compounds evaluated as P-gp inhibitors (amiodarone, captopril, carvedilol, diltiazem, felodipine, isradipine, mibefradil, nicardipine, nifedipine, nitrendipine, quinidine, ranolazine, sertraline, telmisartan, troglitazone, and verapamil) were obtained from Sigma (St. Louis, MO). Other chemicals and reagents were obtained by each participating laboratory from local commercial sources.
Cell Lines and Culture.
Caco-2 cells (ATCC HTB-37) were from the American Type Culture Collection (Rockville, MD) or INSERM U-505 (Paris, France). LLC-PK1 cells expressing cDNAs encoding human MDR1 (LLC-PK1-MDR1 cells) were obtained from the Netherlands Cancer Institute or from BD Gentest (Woburn, MA, cat. no. 450211). MDCKII cells expressing cDNAs encoding human MDR1 (MDCKII-MDR1 cells) were obtained from the Netherlands Cancer Institute, the National Cancer Institute at the US National Institutes of Health (Bethesda, MD), or were generated in-house at the participating laboratory. Table 1–3 indicates which cell line was used by each of the participants, as well as the origin of that cell line.
Culture conditions varied between participating laboratories (Tables 1, 2, 3) since participants were asked to generate IC50 values using the conditions as established and used in their own laboratories. In general, cells were seeded onto permeable filter supports of varying materials and formats, and grown until confluent, polarized monolayers were formed. Some of the laboratories have historical microscopic data to demonstrate formation of cell monolayers (Hidalgo et al., 1989; Tran et al., 2004; Xia et al., 2005). Most laboratories determined the optimal combination of seeding density and time in culture by measuring transepithelial electrical resistance (TEER) and/or paracellular permeability and transport of a P-gp substrate at various seeding densities and times after seeding when they implemented the assay in their company. The conditions were then standardized for subsequent experiments, typically to the earliest time after seeding when each of these controls performs well. The quality controls included in the current study were determination of digoxin bidirectional transport and mass balance (no issue for any of the participants), TEER values, and/or paracellular permeability (the latter two parameters recorded in Tables 1–3).
Membrane Vesicles.
Membrane vesicles prepared from cells expressing P-gp were obtained from Solvo Biotechnology (Budapest, Hungary) following the method described by Keppler et al. (1998) or were generated in-house at the participating laboratory (Table 4) (Karlsson et al., 2010).
Bidirectional Transport Assays across Polarized Cell Monolayers
Assay conditions varied between participating laboratories (Tables 1–3). In general, bidirectional (apical-to-basolateral, or A>B, and basolateral-to-apical, or B>A) transport of the P-gp probe substrate digoxin across polarized Caco-2, LLC-PK1-MDR1, or MDCKII-MDR1 cell monolayers was assessed in the absence and presence of six increasing concentrations of various inhibitors. Digoxin concentrations used in these systems were 13 nM to 10 µM. A positive control inhibitor was included in each assay at a single concentration. Assay buffer containing digoxin and inhibitor (if applicable) was added to the donor compartment, and assay buffer containing inhibitor (if applicable) was added to the receiver compartment. Monolayers were then incubated at 37°C for 45–180 minutes, at which point samples were taken from both donor and receiver compartments. Digoxin concentrations in each sample were determined by scintillation counting or liquid chromatography followed by tandem mass spectrometry (LC-MS/MS).
Vesicle Uptake Assays
Assay conditions varied between participating laboratories (Table 4). Briefly, the ATP-dependent uptake of either 1–2 µM N-methylquinidine or 0.1 µM vinblastine into inside-out membrane vesicles was assessed in the absence and presence of seven different concentrations of inhibitors listed above. The reaction mixtures without ATP or in the presence of 5′-adenylylimidodiphosphate (AMP-PNP, a non-hydrolyzable analog of ATP) were used as a negative ATP control (no ATP-dependent transport activity). The reaction mixtures were incubated at 37°C for 2–3 minutes. The reaction was then stopped by addition of ice-cold sucrose buffer and vesicles were collected onto a filter plate. The substrate concentration in each sample was determined by either scintillation counting or LC-MS/MS. Digoxin was not used as a P-gp probe substrate in vesicles due to a low signal-to-background ratio potentially resulting from binding of digoxin to the extracellular domain of the Na+/K+ ATPase (the pharmacological target for digoxin) in right-side-out membrane vesicles.
Data Analysis
Since previous inhibition data were available in the literature, the number of inhibitor concentrations used for the cell-based experiments was limited to six. Two additional data points were provided by the “positive control inhibitor” and the “no inhibitor” controls. For the vesicle-based experiments seven inhibitor concentrations were used. Each participant was provided with an Excel template for standardized data collection and automated calculation of the fraction transport activity remaining as a function of increasing inhibitor concentration using several different published transport activity equations for the cell lines (Table 5) and a single equation for the vesicles. The cell line template also generated an interpolated IC50 value for immediate assessment of the data.
Data points associated with inhibitor concentrations that showed limited solubility in the assay buffer (as evident by visual inspection) and/or toxicity to the cells (as evident by failing a monolayer integrity test such as elevated permeability of a paracellular low permeability marker) were excluded from the analysis.
Probe Substrate Transport.
For the cell monolayer experiments, digoxin transport was expressed as Papp (apparent permeability) and calculated as follows (Tran et al., 2004):(1)where:
Papp is the apparent permeability in cm/s
VR is the volume of the receiver compartment (ml)
CR is the concentration in the receiver solution at the end of the incubation (μM)
CD is the initial concentration in the donor solution (μM)
A is the surface area of insert filter membrane (cm2)
t is the time interval over which permeability is measured (s).
For the vesicle experiments, the ATP-dependent uptake of N-methylquinidine or vinblastine was determined by subtracting probe substrate uptake in incubations without ATP or with AMP-PNP (nonhydrolyzable analog of ATP) from that in the presence of ATP. Probe substrate uptake was expressed as pmol/min per milligram vesicle protein or as ATP-dependent uptake clearance (µl/min per milligram).
Inhibition of Probe Substrate Transport.
For cell monolayer data, the fraction of transport activity remaining in the presence of inhibitor was initially determined using six different transport activity equations discussed in the recent literature (Table 5, eqs. A1, B1, B2, C1, C2, and D). The final analysis was performed using eqs. A1, B1, C1, and D only. For membrane vesicle data, the fraction of transport activity remaining in the presence of inhibitor was determined as the ratio of ATP-dependent uptake in the presence of inhibitor divided by that in the absence of inhibitor.
IC50 Determinations.
IC50 calculations were not performed if a compound did not inhibit probe substrate transport, or if inhibition did not exceed 50% at the highest inhibitor concentration tested. As a first analysis step, IC50 values for cell monolayer data were determined by linear interpolation in Excel as follows:(2.1)where:
y1 is relative P-gp inhibition calculated by a particular equation when the inhibitor concentration is x1 (µM)
y2 is relative P-gp inhibition calculated by a particular equation when the inhibitor concentration is x2 (µM)
and y1 > 0.5 > y2 and chosen to be closest to 0.5.
This linear interpolation control was put in place to have an easy check on the IC50 values obtained by nonlinear regressions submitted by each of the participants.
Nonlinear regression analysis was performed by participants using various commercially available software packages, including Grafit (Erithacus Software Ltd., East Grinstead, UK), GraphPad Prism (GraphPad Software Inc., La Jolla, CA), Origin (OriginLabs Co, Northampton, MA), Sigma Plot (SPSS Inc., Chicago, IL), WinNonlin (Pharsight Corporation, Raleigh-Durham, NC), Kaleida Graph (Synergy Software, Reading, PA), and XLfit (ID Business Solutions Ltd., Guildford, UK). All data (fraction of transport activity remaining in the presence of inhibitor) were fitted to the standard IC50 equation or Hill equation:(2.2)where:
A([I]) is the activity measured when the inhibitor concentration is [I]
A(∞) is remaining transport when P-gp is completely inhibited by the positive control inhibitor, e.g., GF120918 or in membrane vesicles represents uptake activity in the absence of ATP or presence of AMP-PNP.
A(0) is the activity measured in absence of inhibitor (the negative control)
IC50 is the inhibitor concentration at which A([IC50]) = [A(0) + A(∞)]/2
β is slope factor or Hill coefficient.
For a 4-parameter fit, IC50, β, A(0), and A(∞) were independently and simultaneously fitted. For a 3-parameter fit, the positive control A(∞) was fixed and IC50, β, A(0) were independently and simultaneously fitted. For a 2-parameter fit A(0) and A(∞) were fixed and IC50 and β were independently and simultaneously fitted.
Statistical Analysis
Variance Component Analysis of Digoxin Transport in the Absence of Inhibitor.
The purpose of this analysis was to evaluate variability between cell lines and laboratories with respect to transport of digoxin in the A>B and B>A directions, as well as digoxin efflux ratios. For the variance component analysis of digoxin transport, an analysis was performed using cell lines and laboratories as random effects. Each of the experiments, which encompassed differences in experimental protocols (Table 1–3), was included as a fixed effect. In this analysis, random effects for interactions between cell lines by experiment and laboratories by experiment were also investigated. Replicates, which included well-to-well variability and measurement variability, were used to determine the overall error in the model. Each response was evaluated using this model separately.
Evaluation of the Quality of the IC50 Fit.
Data submitted by participants were of variable quality. To ensure that only robust IC50 values were used for further analysis, an objective statistical test designated as the “t-statistic” was developed to select data sets as follows. IC50 values for all B>A and A>B transport inhibition data and vesicle uptake inhibition data were fitted by one of the participants (without prior data transformation to calculate remaining transport activity, Table 5) to a different version of eq. 2.2, namely eq. 2.3, using 2-, 3-, and 4-parameter fits.
(2.3)The slope factor β is defined by the slope of the inhibition curve at the IC50 and α is defined as the IC50 “locator”. We show this version of eq. 2.2 because the t-statistic developed to score the quality of the fitted data utilizes β, α, and their corresponding SEs.
All replicates were used as input data, rather than just the average transport at each inhibitor concentration. Each company’s positive and negative control data (e.g., transport in the presence of a prototypical P-gp inhibitor to define maximal inhibition, and transport in the absence of inhibitor to define no inhibition, respectively) were incorporated as data points in the fits. The incorporation of the positive and negative controls into the fits was performed because a significant number of inhibition plots lacked well defined upper and lower plateaus (i.e., sigmoidicity), using only the six or seven inhibitor concentrations without the positive and negative controls. Nearly all of the data used inhibitor concentration ranges with roughly 3-fold differences between concentrations. Empirically, it was found that placing the negative control at 3-fold below the lowest inhibitor concentration and the positive control at 9-fold above the highest inhibitor concentration gave fits that did not bias the IC50 values fitted, i.e., the controls did not exert undue influence over the IC50 estimates. If the negative or positive control was placed too close to the nearest inhibitor concentration (<3-fold lower or 9-fold higher, respectively), the fitting routine did not recognize the controls as defining the appropriate plateau. If the negative or positive control was placed too far from the nearest inhibitor concentration (>3-fold lower or 9-fold higher, respectively), the fitted parameters depended strongly on the exact placement of the negative and positive controls. Thus, the placement of the negative and positive controls, if not done properly, could influence the IC50 fits of those data sets lacking an inhibitor concentration defined plateau. Further justification for this empirical treatment was provided by a core panel of 10 participants (described below, in this same section). This panel did not include the single participant who performed all the fits described in this paragraph.
The IC50 is defined by both α and β (eq. 2.3), while the Hill coefficient is defined by β. The SE of α and β will depend on the spread in the data as well as the number of data points on the upper plateau (no inhibition), lower plateau (complete inhibition), and on the slope. The statistical analysis was developed based on the Student t-statistic for each fitted parameter (α and β) (eqs. 3.1 and 3.2):
(3.1)(3.2)The SE and the t-statistic are inversely related; the smaller the SE, the larger the t-statistic for α and β, which corresponds to a better fit. Since the goodness of the IC50 fit depends on both α and β terms, a joint statistic is defined as their geometric average
(3.3)During development of the t-statistic it was found that the optimal statistical analysis was based on the combined t-statistics of a 2- and 3-parameter fit of the same data set (eq. 2.3). The respective t-statistics are denoted as tαβ(2p) or tαβ(3p). A joint combined t-statistic for the 2- and 3-parameter fits was defined as:
(3.4)The t-statistic (tαβ(2p), tαβ(3p) or tαβ(2p*3p) is a descriptor of the quality of the curve fit; that is, the more sigmoidal the curve (i.e., clearly defined plateaus) and the smaller the standard error between replicates, the larger the t-statistic, and correspondingly, the better the fit. Since there was no clearly determined statistically significant “P value” for this novel statistic, a cut off value was established by group consensus as follows: statistical tαβ values for the 2- and 3-parameter fits, denoted as tαβ(2p) and tαβ(3p), respectively (eq. 3.3), as well as the combined statistic for the 2- and 3-parameter fits denoted as tαβ(2p*3p) (eq. 3.4) were generated for each fit. A core panel of ten participants evaluated the amassed best-fit IC50 curves and identified a cut-off value of tαβ(2p*3p) >3 for datasets of sufficient quality for final evaluation. Data sets for which the tαβ(2p*3p) <3 were excluded from further analysis. This panel evaluation and calibration of the t-statistic justifies the empirical treatment of inclusion and placement of negative and positive controls at one-third of the lowest inhibitor concentration and at nine times the highest inhibitor concentration, respectively. In cases where this placement of the controls was not in accord with the six or seven inhibitor concentrations and the controls exerted undue influence over the IC50 fits, the resulting t-statistic is low (poor quality fit).
After this t-statistic was derived and used in Matlab (MATLAB Version 7.11.0.584 (R2010b), we found that obtaining α and SEα from the common fitting software packages was difficult. Obtaining SEβ was easy and direct, since β is the slope factor. Subsequently it was shown that a t-statistic defined as(3.5)gave essentially the same values as tαβ(2p3p) and can be used to assess data quality. Here β is used as a surrogate for α. The terms tββ(2p) and tββ(3p) are surrogates for tαβ(2p) and tαβ(3p) in eq. 3.4. We found only a few cases in all of the curated data where t αβ(2p*3p) < 3 < t ββ(2p*3p).
The t-statistic based on SEβ is straightforward to implement. It is easily available in common software packages used for logistic regression and is estimated similarly in different software packages, unlike the standard error of the IC50 estimate. Additionally tβ (as well as tα) should be (nearly) unaffected by data transformations including those used to perform 2-, 3-, and 4-parameter logistic fits—unlike the root-mean-square error (RMSE) and other measures of residual error. The t-statistics are strongly, though not necessarily linearly, related to other measures of fit quality such as the standard errors of α, β, and log(IC50), and measures of residual error such as RMSE (root-mean-square error) and r2. A detailed analysis of this relationship will be presented in a future publication (O’Connor et al., personal communication).
There were several cases where the IC50 ≈ 1 µM, in which case ln{IC50} ≈ 0, which caused the t-statistic to be artifactually small. This was alleviated by fitting the data using units of molar (M), rather than µM.
Variance Component Model Analysis of IC50 Values.
The variance component analysis represents a nested design, in which the laboratories are completely nested within the systems (that is each laboratory used only one single P-gp expression system). Furthermore, each laboratory contributed a single IC50 value, but for multiple inhibitors. For the variance component analysis of the IC50 data, an analysis was performed using cell lines and laboratories as random effects and experiment (each time an IC50 value was determined) as a fixed effect. This is a commonly used analysis for determining the components of variability. In this analysis, experiment encompassed differences in the experimental protocols, as well as the additional variability due to the use of multiple inhibitors. For the IC50 data, only a single value was calculated and reported for each cell, laboratory, and inhibitor combination. No further tests could be done on interactions. Replicates included random effects due to cell-by-inhibitor interactions, laboratory-by-inhibitor interactions, well-to-well variability, and measurement variability. Each response was evaluated using this model separately. Values below 1 and above 1000 were omitted from the analysis, and the IC50 values were logged using a natural log transformation as the raw data did not appear normally distributed. A similar analysis was also performed to determine the source of variability when considering all systems together, cell lines, and vesicles.
Principal Component Analysis.
Principal component analysis (PCA) identifies patterns in complex data sets and expresses the data highlighting similarities and differences in the underlying components. The PCA is like a least squares fit in multiple dimensions. In a least squares regression the straight line fitted is chosen to maximally explain the variance of the data. Similarly, the PCA chooses the first axis through a “cloud” of data points in the direction that explains the largest variance in the data. The second axis is orthogonal to the first and chosen to explain most of the residual variance in the data. For the robust PCA analysis a complete set of IC50 values is needed from each participant. Therefore, unlike the variance component analysis, the PCA analysis was limited to IC50 values from eight inhibitors for a total of 24 data sets, using cell line-based as well as vesicle-based IC50 values. PCA analysis was performed using the algorithm for PCA via singular value decomposition (SVD) provided by Press et al. (2007) and run in Matlab. The results were compared against PCA performed in standard programs like SAS and R with identical results.
Results
The process of data generation and analysis by the “P-gp inhibition working group” is illustrated in Fig. 1. Experimental data for P-gp inhibition by 16 compounds reported to be inhibitors of P-gp, were generated by all participating laboratories using their respective established methods (Tables 1–4). In the cell-based experiments, each participant used digoxin as the probe substrate, whereas N-methylquinidine or vinblastine was used in the vesicle-based experiments. The inhibition experiments were performed in a protein-free assay buffer. The primary difference in methodology between the laboratories was the P-gp expression system used, i.e., three different cell lines and inverted membrane vesicles. There were also differences in other experimental conditions: 1) those potentially affecting the properties of the cells, such as cell line origin (in case of the MDCKII-MDR1 and LLC-PK1-MDR1 cells), cell maintenance procedures, cell passage number, seeding density, growth media, and time in culture, or the properties of the vesicles, such as cell source and procedure for vesicle preparation; and 2) those conditions applying to the inhibition experiment, such as assay buffer, probe substrate and probe substrate concentration, pre-incubation with inhibitor, and plate shaking. Since captopril did not inhibit P-gp in any of the systems used it was not included in the analysis of P-gp IC50 variability.
Probe substrate transport data (Papp values for cell models and uptake values for vesicles) were entered into a standardized Excel spreadsheet to calculate fraction P-gp activity remaining as a function of inhibitor concentration. The use of a standardized Excel spreadsheet to capture raw data proved invaluable in the data checking and analysis process. The P-gp activity in the respective cell lines was calculated using various equations from the literature that were incorporated in the spreadsheet (Table 5). Subsequently, the inhibition data were subjected to a statistical test (t-statistics) to ensure data were sufficiently robust for IC50 fitting. Final IC50 values were then determined using nonlinear regression by each of the participants for each of the transport activity equations. The standardized Excel sheet also estimated an initial IC50 value for each transport activity equation by linear interpolation of the inhibition data, providing a means to verify the fitted IC50 value derived from each of the transport activity equations. Detailed results of each step are outlined below.
Transport of Probe Substrates.
Table 6 summarizes digoxin Papp values in the B>A and A>B transport direction, in the absence and presence of a positive control P-gp inhibitor, across the three cell lines in the various participating laboratories. Donor chamber digoxin concentrations varied between participants, from 13 nM to 10 µM. Initial experiments performed by several participants demonstrated that the Papp value for digoxin was independent of donor concentration in the range used here (unpublished data). Substantial interlaboratory variability was observed in the Caco-2-derived data. Variability in TEER values, paracellular permeability, and transporter expression in Caco-2 cells has been noted previously (Walter and Kissel, 1994, 1995) and has been ascribed to cell source, passage number, and culture conditions (Behrens and Kissel, 2003; Hayeshi et al., 2008; Volpe, 2008; Miliotis et al., 2011). Less variability in digoxin transport was observed across MDCKII-MDR1, and even less with the LLC-PK1-MDR1 cell lines. There were also fewer participants in the MDCKII-MDR1 and LLC-PK1-MDR1 groups compared with the Caco-2 group. Variance component analysis using cell lines and laboratories as random effects and each experiment measuring digoxin transport as fixed effect shows that the largest source of variability in A>B and B>A transport is the laboratory-to-laboratory variability (∼80%), rather than a difference between the three cell lines (≤9%) (Table 7). This is true for efflux ratios as well, with 62% of the variance ascribed to laboratory-to-laboratory variability and ≤8% to a difference between the cell lines (Table 7). Table 8 summarizes uptake of the P-gp probe substrates N-methyl quinidine or vinblastine into P-gp-expressing vesicles in the presence and absence of ATP. The vesicles also show substantial variability in uptake of probe substrate. This may be due in part to differences in P-gp expression systems used for vesicle preparation (e.g., differences in P-gp expression levels, membrane composition) or differences in the proportion of inside-out to right-side-out vesicles. It is also possible that differences in probe substrates (N-methylquinidine versus vinblastine) and differences in postincubation vesicle collection and washing procedures contribute to the observed variability in ATP-dependent uptake clearance, as may differences in the thawing and handling of the vesicles prior to the assay.
Transport Inhibition.
The purpose of the current study was to investigate whether this intrinsic interlaboratory difference in transport of probe substrates also results in an inter-laboratory difference in inhibition of transport, expressed as the IC50 values. The 16 inhibitors used in this study were selected based on the availability of robust clinical DDI data with digoxin (Fenner et al., 2009; Cook et al., 2010) and are listed in the Materials and Methods section.
Data Quality.
Initial visual inspection of the transport inhibition curves (∼1300) generated by participants (facilitated by the standardized spreadsheets) revealed data quality aspects that needed to be addressed before further data analysis. Two common observations were 1) the “no inhibition” plateau on the left of the inhibition curve was not consistently well defined as illustrated by the transport inhibition of digoxin across Caco-2 cells by carvedilol and troglitazone in Fig. 2, A and B, respectively and 2) the “complete inhibition” plateau on the right of the inhibition curve was defined often by only the positive control inhibitor as illustrated by transport inhibition across LLC-PK1-MDR1 cells by ranolazine and across MDCKII-MDR1 cells by verapamil in Fig. 2, C and D, respectively. The data point at the extreme left of the curve represents the digoxin Papp value in the absence of inhibitor, while the data point at the extreme right represents the digoxin Papp value in the presence of the positive control inhibitor (see Materials and Methods). Another issue sometimes encountered was substantial variability in the triplicate data points as shown in Fig. 2, B and C. For these reasons, the authors implemented an objective statistical analysis (t-statistic), based on 2 and 3-parameter fits to the data using a logistic equation, eq. 2.3, that identified data sets that gave robust IC50 values, as described in the Materials and Methods section. The t-statistics for the data shown in Fig. 2 will be discussed below, after a description of the implementation of this statistical analysis.
When scrutinizing the resulting fits to eq. 2.3, the authors found that the same primary data could result in the same or different IC50 values depending on the data quality and fitting algorithm, i.e., a 2-, 3 or 4-parameter fit. An example of this is provided in Fig. 3. Two and four parameter fits (eq. 2.3) of three representative data sets of acceptable quality are shown. (A) and (B) show the inhibition of digoxin transport through MDCKII-MDR1 cells by verapamil. While the no inhibition plateau on the left is well defined and there are several points along the slope, the complete inhibition plateau is only defined by the positive control inhibitor itself. However, since the positive inhibitor control value was in accord with all the other points, the IC50 values of the 2- and 4-parameter fits are quite close. (C) and (D) show the inhibition of digoxin transport in Caco-2 cells by diltiazem. This curve (and others for diltiazem) was very broad, suggesting that inhibition does not occur at either a single-binding site or a single transporter, in contrast to, for example, verapamil (Fig. 3, A and B). However, also for diltiazem, the IC50 values of the 2 and 4-parameter fits are quite close (Fig. 3, C and D). (E) and (F) show the inhibition of vinblastine uptake in vesicles by quinidine. The data are nearly a straight line and the IC50 values generated by a 2 and 4-parameter fit are about 2-fold different due to the uncertainty in the plateaus and absence of sigmoidicity.
To avoid introducing variability associated with the fitting algorithm, a 2-parameter fit was applied to all data as the upper and lower plateaus were constrained to no inhibition and complete inhibition. Since in this case upper and lower plateaus are defined rather than fitted, the resulting fits are more constrained and therefore less susceptible to variation.
Quality of the IC50 Curves (Implementation of t-Statistic).
To objectively determine which data sets yielded robust data for final data analysis, a statistical metric (known as the t-statistic) was developed as described in the Materials and Methods section. The t-statistics (tα, tβ, tαβ) (eqs. 3.1, 3.2, 3.3, 3.4) were used to assess the sigmoidicity of the IC50 curve fit. The statistics on α and β were selected because these coefficients were directly estimated as part of the logistic regression analysis, and were used to calculate the derived estimate of IC50 values. A t-statistic tαβ(2p*3p) <3 cut-off was selected for exclusion of data sets.
Some examples of the application of the t-statistics are shown using Figs. 2 and 3. Fig. 2A shows the inhibition of B>A digoxin transport across Caco-2 cells by increasing concentrations of carvedilol. The t-statistic tαβ(2p*3p) for this data set is 2 and therefore these data are excluded from further analysis. For the 2 and 3-parameter fits of these data the individual t-statistics were tαβ(2p) = 5.7 and tαβ(3p) = 0.9. The 2-parameter fit was more tolerant of the poorly defined no-inhibitor plateau on the left than the 3-parameter fit, resulting in the poor combined t-statistic. Since in a 2-parameter fit the no-inhibitor and complete-inhibition plateaus are fixed, the t-statistic for a 2-parameter fit will be equal to or greater than that for a 3-parameter fit.
As described in the Materials and Methods section, after data analysis was completed a slightly different t-statistic, tββ(2p*3p), was tested. The later statistic is more easily obtained from commonly used software packages and therefore of greater general utility. It was found that there were a few cases in our data set where tαβ(2p*3p) < 3 < tββ(2p*3p). One of these is the carvedilol example shown in Fig. 2A. Here the tββ(2p*3p) = 4.1, while tαβ(2p*3p) = 2.
Fig. 2B shows inhibition of B>A digoxin transport by troglitazone across Caco-2 cells. The combined tαβ(2p*3p) = 3.7 and tββ(2p*3p) = 3.7 results in the inclusion of this data set for IC50 variability assessment. Fig. 2C shows inhibition of B>A digoxin transport by ranolazine (LLC-PK1-MDR1 cells). The combined tαβ (2p*3p) = 5.0 and tββ (2p*3p) = 4.9 results in acceptance of this data set. Fig. 2D shows inhibition of B>A digoxin transport by verapamil (MDCKII-MDR1 cells). The combined tαβ (2p*3p) = 18.6 and tββ (2p*3p) = 18.6 was one of the highest found for the cell data, despite the fact that the complete-inhibition plateau on the right was defined by the positive control inhibitor. This was because the replicates were very close and the well-described no inhibition plateau on the left largely defined the IC50 value. The amplitude of the positive control inhibitor on the right agreed with the rest of the fit.
Fig. 3 illustrates the practical meaning of the t-statistic cut-off for verapamil as measured by two laboratories in MDCK II-MDR1cells (Figs. 3, A and B), for diltiazem measured across Caco-2 cells (Figs. 3, C and D) and for quinidine using the vesicle assay (Figs. 3, E and F).
Fig. 3, A and B data are of high quality, as judged by tαβ(2p*3p) = 18.6 and tββ(2p*3p) = 18.6. The IC50 values for both 2- and 4-parameter fits are essentially identical. The data shown in Fig. 3, C and D are acceptable with tαβ(2p*3p) = 5.5 and tββ(2p*3p) = 5.4. The IC50 values for both fits are about 1.7-fold different. The data shown in Fig. 3, E and F are also acceptable with tαβ(2p*3p) = 4.9 and tββ(2p*3p) = 4.2. The IC50 values for both fits are about 1.8-fold different, but there is a 5.8 μM difference in the absolute values. The individual statistics for the 2 and 4-parameter fits were tβ(2p) = 6.2 and tβ(4p) = 0.2, showing lesser confidence in the 4-parameter fit accounting for the different estimates of the IC50 values. These examples illustrate the utility of the t-statistic. The larger its value above 3, the smaller the probability that the fitted IC50 value is significantly wrong. While the data quality was assessed by the combined t-statistic on a 2- and 3-parameter fit, all of the subsequent data analysis of the curated data was performed using the IC50 values obtained by a 2-parameter fit, as mentioned above.
Table 9 summarizes the result of the statistical analysis of the data quality. Overall, the data quality obtained with the vesicles was higher than with the cells. For the cells, the B>A data are more robust than the A>B data, which is probably related to the relatively low transport of digoxin in this direction and the higher associated variability.
Variability in IC50 Values.
The IC50 values for each of the fifteen compounds in each of the four experimental systems are shown in Fig. 4 (see Supplemental Data for individual IC50 values). As mentioned above, one of the 16 compounds was not included in the variability analysis (captopril) because it did not inhibit P-gp in any of the systems used. For the cell line-derived data, IC50 values based on remaining transport activity eqs. A1, B1, C1 and D are included in this figure. Since a substantial fraction of the data were excluded due to unacceptable tαβ(2p*3p) values, the number of IC50 values for each compound, system and transport activity equation in Fig. 4 is not the same. For the efflux ratio based eqs. A1 and B1, IC50 values were included in the analysis only if A>B as well as B>A data were of acceptable quality (tαβ(2p*3p) > 3). The reason for this is the substantial influence of the A>B data on the efflux ratio. For equation D, IC50 values were included if B>A data were of acceptable quality, without consideration of the quality of the A>B data. The reason for this is that the effect of the A>B data in equation D is minimal (subtraction of a small number from a large number) when digoxin is the probe substrate (see also section on equation comparison below).
As mentioned above, participants used digoxin concentrations between 13 nM and 10 µM for the cell-based experiments. One of the participants determined IC50 values in MDCKII-MDR1 cells for GF120918 and verapamil at digoxin concentrations of 30 nM, 500 nM and 10 µM and found no difference in the resulting IC50 values (data not shown).
The lowest variability in IC50 values was observed for isradipine (Inhibitor 5) and sertraline (Inhibitor 12) (24- and 20-fold difference between lowest and highest IC50 values) and the highest for verapamil (Inhibitor 15) and telmisartan (Inhibitor 13) (796- and 407-fold range). In these two extreme cases, the variability was greatly influenced by data from one contributor, but a different one in each case. For telmisartan, a very similar value was reproduced by the company that generated the extremely low IC50 value. Ignoring these two extreme data points brings the verapamil and telmisartan variability down to 159- and 69-fold, respectively.
Transport Activity Equation Comparison.
Fig. 5A shows that transport activity eqs. A1 and A2 (efflux ratio-based equations) are essentially the same. Equation A1 includes a positive and negative control value in the calculation of transport activity, whereas eq. A2 does not. Equation A2 is the FDA recommended equation (US FDA/CDER, 2006, 2012). For this particular initiative, due to the limited number of inhibitor concentrations in the cell line experiments, the participants were asked to fit data to eq. A1 so that a no inhibitor control and positive control inhibitor could be included as data points (see Materials and Methods section). Note in Fig. 4 that the lowest cell-based IC50 values for all compounds are based on transport activity eqs. A1 (blue diamonds) and C1 (red triangles) and the highest on eqs. B1 (green squares) and D (black circles), with the exception of Inhibitor 15 (verapamil). It has been noted previously for MDCKII-MDR1 cells as well as LLC-PK1-MDR1 cells (Lumen et al., 2010, Sugimoto et al., 2011) that an IC50 based on transport activity expressed as an efflux ratio (eq. A1 or A2) is typically lower than one based on B>A transport (B1). It is shown in Fig. 5, B and C that transport activity eqs. A1 and C1 result in very similar IC50 values, as do eqs. B1 and D. By contrast, eqs. B1 and D typically give higher values than eqs. A1 and C1 (Fig. 5D, only comparison shown is for eqs. A1 and B1). Equation B1 typically results in an IC50 value which is on average 3-fold higher than the IC50 value resulting from eq. A1. In 23% of the cases eq. B1 generates an IC50 value which is at least 6-fold higher than the eq. A1 IC50 value. Therefore, some of the variability in IC50 values shown in Fig. 4 is due to the use of different equations to calculate remaining transport activity. For example the highest nicardipine cell-based IC50 value (based on eqs. A1, B1, C1 and D combined) was 46 and the lowest 0.3 µM, resulting in a 148-fold difference. Considering only eq. B1 the highest and lowest IC50 values were 21.4 and 0.8 µM, a 27-fold difference. Another example is quinidine, with highest and lowest cell-based IC50 values (based on eqs. A1, B1, C1 and D) of 78 and 0.8 µM (98-fold difference), whereas eq. B1 resulted in highest and lowest IC50 values of 73 and 2.2 µM, a 33-fold difference. Taking vesicle IC50 values into account as well as eq. A1, B1, C1 and D-based IC50 values, the least variability was observed for sertraline, with a 20-fold difference between lowest and highest IC50 values and the greatest variability with verapamil (796-fold difference). Considering vesicle IC50 values and only eq. B1-based IC50 values, the least variability of 6-fold was observed with troglitazone and the greatest of 406-fold with telmisartan. Hence, standardizing the cell-based IC50 determination to one transport activity eq. (B1) reduces the lowestfold variability across cell- and vesicle-based assays from 20-fold (sertraline) to 6-fold (troglitazone) and the greatestfold variability from 796-fold (verapamil) to 406-fold (telmisartan). However, substantial variability remains.
Equation B1 (based on B>A transport) declines in magnitude as inhibitor is added causing the B>A([I]) flux to decline. As inhibitor is added, eq. A1 (based on efflux ratio) declines in magnitude as the B>A([I]) flux declines in the numerator, like eq. B1, but additionally declines as the A>B([I]) flux increases in the denominator. Therefore eq. A1 is declining faster than eq. B1, resulting in a lower IC50 value. The smaller IC50 generated by eq. A1 relative to the value resulting from eq. B1 is simply algebraic. Equation B1 (based on B>A transport) and D (based on net flux: B>A minus A>B transport) result in similar IC50 values. This may not be the case for all P-gp probe substrates, but for the probe substrate digoxin this is likely due to the fact that the absolute decline of digoxin transport in the B>A direction is greater than the absolute increase in transport in the A>B direction for any given inhibitor concentration (exemplified in Table 6 for the positive control inhibitor). This minimizes the effect of the increase in A>B transport on the decline in the B>A direction and thus on the IC50 value obtained for the net flux equation.
For further statistical analyses (variance component analyses and principal component analysis) we used IC50 values obtained without prior data transformation as in the transport activity eqs. A1, B1, C1 and D. This data transformation results in propagation of error (when subtracting or dividing two numbers, each with associated error), which can be avoided by simply applying the IC50 fitting algorithm to untransformed data (Papp values in presence of inhibitor with positive control inhibitor and no inhibition controls included as data points in the fits), as described in the materials and methods section under “Evaluation of the Quality of the IC50 Fit.” This particular data treatment and IC50 fitting protocol results in the largest, most robust IC50 data set. The resulting IC50 values will be most similar to those based on unidirectional flux transport activity equations, the only difference being data normalization prior to IC50 fitting. Standardizing the cell-based IC50 determination to eq. 2.3 applied to untransformed data reduces the lowestfold variability across cell- and vesicle-based assays from 20-fold (sertraline) to 7-fold (troglitazone) and the greatestfold variability from 796-fold (verapamil) to 354-fold (telmisartan).
Variance Component Analysis.
The variance component analysis represents a nested design, where the laboratories are completely nested within the systems. Experimental system and laboratory were used as random effects. The effect of inhibitor was accounted for as a fixed effect. Each laboratory was asked to complete testing on all inhibitors. This is a commonly used analysis for determining the components of variability. In Figs. 6, A and B, the large blue rectangles represent the variability across systems and the small red rectangles represent variability across laboratories. Therefore, it is easy to see that systems, especially the cell-based systems contribute little variability, whereas laboratory-to-laboratory variability, even within the same system, is higher.
For cell-based IC50 values obtained for transport inhibition in both A>B and B>A directions the largest source of variability was interlaboratory differences, 35% in either direction (Table 7 and Fig. 6, A and B). The variability due to the use of different cell lines is substantially smaller at 13% and 14%, respectively. The remaining variability (51–52%) was due to replication, which in this case constitutes interactions between experiment and cell, experiment and laboratory, well-to-well variability, and measurement variability, and could not be further deconvoluted because only one IC50 value is reported by each company for each inhibitor. In addition to the laboratory-to-laboratory variability in absolute IC50 values, the rank-ordering of inhibitor potency also varied from laboratory to laboratory (Fig. 6, A and B). When vesicle- and cell-based (B>A only) IC50 values were combined, variance component analysis demonstrated that laboratory-to-laboratory variability remained high at 29% and was approximately equal to system variability at 28%, while variability due to replication was 43% (Table 7 and Fig. 6B).
Principal Component Analysis.
As mentioned in the Materials and Methods section, for a robust PCA analysis a complete set of IC50 values is needed from each participant. Consequently, the PCA analysis performed here did not include all inhibitors. Furthermore, the PCA analysis was only performed for IC50 values based on inhibition of B>A transport. Table 10 shows axes 1–8 for the PCA performed on eight compounds evaluated in four systems (Caco-2, MDCKII-MDR1, LLC-PK1-MDR1 cells, and vesicles) with IC50 values contributed from all but two laboratories. Axis 1 explains 54% of the variance in the data, axis 2 explains 14%, and axis 3 explains 13%, i.e., the first three axes explain greater than 80% of the variance. The remaining axes each explain <10% of the variance in the data. Notably in axis 1, the weights (i.e., the values of the eigenvectors) of the eight drugs are all approximately 0.3, indicating that the bulk of the variance is due to the difference in the average log10{IC50} value for the eight compounds generated by each participant in one of the four systems. For axis 2, most of the weights are close to 0, apart from the maximum weight of 0.58 for mibefradil and the minimum weight of –0.70 for nicardipine. For axis 3, most of the weights are close to 0, apart from the maximum weight of 0.5 for mibefradil and the minimum weight of –0.78 for diltiazem. This indicates that the second and third largest variability observed are derived from the IC50 values for mibefradil, nicardipine, and diltiazem, suggesting that these compounds interact quite differently in some systems.
Basically, the PCA created a two-dimensional map of the “relatedness” between the cells according to their aggregate IC50 values for the eight inhibitors. Fig. 7 shows the graphical display of the PCA for axis 1 and 2. Of the four systems, the MDCKII-MDR1 and LLC-PK1-MDR1 cell lines are closely related with respect to the average log10{IC50} values for the eight inhibitors multiplied by their Eigen values (Table 9), i.e., axis 1. The Caco-2 cell lines varied much more widely, with some Caco-2 laboratories having average log10{IC50} values similar to the MDCKII-MDR1 and LLC-PK1-MDR1 cells and other laboratories having much lower average log10{IC50}values. The vesicles ranged broadly with the Caco-2 cells, with the highest average log10{IC50} value just below the smallest average log10{IC50} value for the MDCKII-MDR1 and LLC-PK1-MDR1 cell lines. While the variance component analysis quantifies the overall variation in IC50 values, it does not show this “relatedness” between the various systems with respect to their IC50 values. To gain further insight into the reason for the variation, the PCA points out which systems and which inhibitors to examine more closely.
Discussion
Given the importance of P-gp in the absorption, disposition, and excretion of drugs, it is essential to have in vitro assays that assess potential DDIs with P-gp during drug development. Currently, a variety of biologic systems are used to assess the potential for these DDIs, including membrane vesicles and nonpolarized and polarized cells expressing P-gp (either endogenously or by expression of transfected constructs encoding P-gp). The International Transporter Consortium (ITC) publication (International Transporter Consortium et al., 2010) and FDA’s 2012 draft Drug Interaction Guidance suggested the use of polarized cells, such as Caco-2 or P-gp transfected cell lines (e.g., MDCKII-MDR1), for the assessment of inhibitor interaction potential and established cut-off values for identifying when a clinical DDI study for P-gp is warranted. However, it has not been established whether different biologic systems, or even biologic systems of similar type, such as polarized cell lines, yield similar IC50 values across different laboratories. Additionally, given the abundance of transport activity equations used in the literature to determine an IC50 value for P-gp, there has been little apparent consensus regarding how to calculate an IC50 value for efflux transporters expressed in polarized cell systems and whether the existing equations for doing so generate similar IC50 values. The potential concern created by the current decision trees for P-gp is that the variability of IC50 values can impact the DDI risk assessment, i.e., some biologic systems and equations may suitably identify molecules that necessitate an in vivo DDI study, whereas other systems may either over- or underestimate the need for conducting such a study, generating an undesirable number of false positives or false negatives. This study systemically evaluated both the biologic systems and the transport activity equations used for the in vitro determination of the IC50 values against P-gp through collaborative work among 23 companies and an academic institution studying a common set of known P-gp inhibitors. The companion article (Ellens et al., 2013) describes the derivation of new decision criteria based on a receiver operating characteristic analysis using the IC50 values described in this article. The companion article (Ellens et al., 2013) also describes how to incorporate P-gp IC50 variability in the DDI risk assessment to optimize decision making.
Substantial interlaboratory variability was observed in the probe substrate transport data as well as the transport inhibition data (IC50 values) across all systems and equations. Caco-2 cells are derived from a human colon carcinoma and are known to express many transporters in addition to P-gp (Lowes et al., 2003, Hilgendorf et al., 2007; Hayeshi et al., 2008). MDCKII-MDR1 and LLC-PK1-MDR1 cell lines are derived from canine and porcine kidney, respectively. Limited information is available on transporter expression in the latter two (for example Di et al., 2011). We are not aware of studies comparing expression levels of multiple transporters across all three cell lines.
The least variability in IC50 values submitted by the participants was seen with sertraline and the greatest with verapamil, with a 20- and 796-fold difference between lowest and highest IC50 values, respectively, clearly indicating the need for standardization, as suggested by Zhang et al. (2008). Standardizing the cell-based IC50 determination to eq. 2.3 applied to untransformed data reduces the lowestfold variability across cell- and vesicle-based assays from 20-fold (sertraline) to 7-fold (troglitazone) and the greatestfold variability from 796-fold (verapamil) to 354-fold (telmisartan). This still leaves substantial variability across cell- and vesicle-based IC50 values, with 13 of the 15 inhibitors showing a difference between lowest and highest IC50 values of at least 20-fold, and only two inhibitors showing less than a 20-fold difference.
The variance component analysis of the variability in probe substrate transport and IC50 values shows that much of the variance observed was due to laboratory-to-laboratory differences rather than due to implicit systematic differences between the P-gp expression systems. The PCA analysis is consistent with this observation and, more importantly, provides insight into the relationship between the various systems with respect to their aggregate IC50 values, as well as into which inhibitors contribute most to the observed variability. This insight will help guide further investigations into the reasons for the variability.
In the cell systems only, the laboratory-to-laboratory differences constitute differences in properties of the cells as maintained in each laboratory, differences in the monolayers on the Transwell inserts as prepared in each laboratory, and differences in experimental protocols to measure inhibition. When vesicles are added to the analysis, differences in the expression systems used for vesicle preparation and vesicle preparation procedures, as well as differences in inhibition study protocols, contribute to the laboratory-to-laboratory differences, as well as the overall variability. Each of the factors and their potential contribution to variability are discussed below.
The IC50 value for P-gp in confluent monolayers is a complicated function of substrate and inhibitor efflux and influx kinetics (Lumen et al., 2010). For example, consider two cell lines identical in all respects except that the P-gp surface density in cell line 2 is greater than that of cell line 1. In the absence of inhibitor, cell line 2 will have a smaller intracellular probe substrate concentration than cell line 1, so the fraction of P-gp with substrate bound to it in the absence of inhibitor will be smaller in cell line 2. This means that more inhibitor will have to be added to cell line 2 to reduce the fraction of P-gp with substrate bound to it by 50% (Lumen et al., 2010). Thus, the IC50 value for cell line 2 will be higher than the IC50 value for cell line 1. This result is simply a consequence of the fact that the smaller the fraction of substrate bound to any site, the larger the amount of inhibitor required to reduce the bound substrate by 50%. If, on the other hand, cell line 2 only differed from cell line 1 by having a larger substrate passive permeability, with identical P-gp surface densities, then cell line 2 would have a lower IC50 value than cell line 1. Therefore, differences between cell monolayer preparations in surface density of P-gp or passive permeability for digoxin are likely to be important contributors to the observed variability in IC50 values derived from cell lines.
The fact that the MDCKII-MDR1 cell line has a basolateral digoxin uptake transporter that is inhibited by GF120918 (Acharya et al., 2008) adds a separate layer of complexity to the interpretation of an IC50 value for P-gp. Any inhibitor that can bind to this digoxin uptake transporter would decrease digoxin efflux and thereby affect the measured IC50 value such that this value is not a pure P-gp IC50. Hence the presence of digoxin uptake transporters and differences in their expression levels may also contribute to variability in IC50 values. The contribution of several different factors to the value of the IC50 in cell lines might explain not only the laboratory-to-laboratory variability in the absolute value of the IC50, but also the laboratory-to-laboratory differences in inhibitor potency rank-ordering.
In the current study, participants used digoxin concentrations between 13 nM and 10 µM. Simulations using the model published in Agnani et al. (2011) with quinidine and verapamil as inhibitors have shown that the IC50 value in a given system is affected by digoxin concentration, but at digoxin concentrations > 10 µM only (unpublished data). IC50 values determined in two different cell lines with different P-gp surface densities will be strongly dependent on P-gp surface density, but not dependent on digoxin concentrations of ≤10 µM. In the current study only two laboratories used 10 µM digoxin. Hence the differences in digoxin concentrations used in this study are not likely to contribute a great deal to the observed IC50 variability. Moreover, removing data from laboratories using digoxin concentrations ≥5 µM did not affect our conclusion (removing laboratories 5, 8, 15, 18, and 19 from Fig. 7). Similarly, preincubation and plate-shaking conditions did not explain laboratory-to-laboratory differences.
Overall, while this variability analysis did not identify the source of the variability, it appears likely that the substantial variability in cell-based IC50 values is largely due to cell source, passage number and culture conditions, as previously concluded for differences between laboratories in TEER values, paracellular permeability and transporter expression in Caco-2 cells (Hayeshi et al., 2008), rather than to an implicit difference between cell lines. Standardizing cell type, cell source, culture conditions and passage number could reduce interlaboratory variability in cell-based IC50 values for inhibition of digoxin transport. Sambuy et al. (2005) performed an interlaboratory study of variability in TEER, mannitol permeability, and alkaline phosphatase expression in Caco-2 cells with the aim to establish a standardized protocol allowing a meaningful comparison of results obtained in different laboratories. Such a study has not yet been performed for transporter expression. While this might be a very worthwhile effort, it is beyond the scope of the current investigation. Some general recommendations can be made with respect to cell culture practices. When first establishing the assay, it is recommended that each laboratory determines the optimal conditions to ensure properly formed tight monolayers by monitoring TEER values, paracellular permeability and efflux transport as a function of seeding density and time. Existence of monolayers should be confirmed by microscopy. The assay should then be performed soon after formation of properly functioning monolayers. Each time the assay is performed quality controls should be included, such as measurement of TEER values, paracellular permeability, and transport activity.
The variability in uptake of probe substrate and IC50 values obtained with vesicles from the same source was surprising given the relatively straightforward assay setup and format. It is recommended to ensure that commercially obtained vesicles perform according to the quality control data provided by the vendor.
The multitude of transport activity equations used in the estimation of P-gp IC50 values in cell lines is a consequence of the asymmetry of the experimental system, i.e., the ability to access P-gp from two directions, A>B and B>A. For the commonly used probe substrate digoxin, flux in the B>A direction is much greater than that observed in the A>B direction; additionally, there is a substantial transport component not inhibited. P-gp transports substrates from the inner leaflet of the plasma membrane to the outside of the cell. If the probe substrate only interacts with P-gp and if inhibitor is added to both donor and receiver chambers, the IC50 value measured in the B>A direction should be the same as that in the A>B direction. To assess the transport activity of P-gp as a function of inhibitor concentration, there are several equations that incorporate transepithelial flux in both directions. However, the rationale for favoring bidirectional flux activity equations (A1, A2, and D) rather than unidirectional flux activity equations (B1, B2, C1) is typically not provided. Our analysis demonstrates that algebraic differences between equations result in different IC50 values.
The simplest experimental design recommended to determine P-gp IC50 values when using digoxin as probe substrate in confluent cell monolayers is the measurement of B>A flux in the presence of increasing inhibitor concentrations. In addition, eight to 12 inhibitor concentrations of equal spacing (e.g., half-log dilutions) are recommended with a wide enough range to define both the upper and lower plateaus, with at least three points in the linear range of the IC50 curve. However, depending on compound solubility, if the lower plateau cannot be realized, a “true” positive control inhibitor, with complete inhibition of P-gp, must be used to define the plateau. The logistic eq. 2.2 is recommended to fit the IC50 value as data transformation in transport activity equations A1, A2, B1, and D contributes error to the IC50 value and standard statistical analyses are inappropriate (examined in greater detail in a future publication by O’Connor et al.). Furthermore, if sigmoidicity is questionable, the application of the tββ(2p*3p) statistic to judge the quality of the fit is also recommended.
In summary, substantial variability was observed in the participant submitted IC50 values for the 15 compounds examined, for cell lines as well as vesicles. This difference was to a large extent due to interlaboratory differences in behavior of the P-gp expression systems. The IC50 variability likely derives from the properties of the P-gp expression system used in a particular laboratory (e.g., P-gp surface density and involvement of other transporters), even when the same cell line is used in different laboratories. The feasibility of standardizing to a single expression system, with a consistent behavior over time, has not yet been demonstrated. The use of different transport activity equations fractionally contributed to the variability observed with the cell lines.
The authors recommend: 1) using best cell culture practices to establish a tight monolayer of polarized cells, 2) ensuring digoxin concentration is sufficiently below the (apparent) Km for the system such that it does not affect the IC50 value, 3) standardization to a single experimental design (digoxin B>A transport only) and IC50 fitting routine (logistic eq. 2.2 or 2.3 applied to transport inhibition data without prior data transformation). When using commercially obtained vesicles it is recommended these are treated as recommended by the vendor and perform according to the quality control data provided.
Finally, the analysis performed here does not identify a P-gp expression system that outperforms the others, but it does point out that the decision criteria as currently derived may not be optimal. More appropriate DDI risk assessment criteria are presented in the companion article (Ellens et al., 2013).
Acknowledgments
The authors thank Houda Hachad and the UW Drug Interaction Database Program for providing the data-sharing platform (www.druginteractioninfo.org). The authors also thank the following individuals for their expert assistance with the experiments: Malin Forsgard and Karima Ben Tabah (AstraZeneca R&D Mölndal), Lawrence Gan (Biogen-Idec), John Herbst, Janet Kolb, and Anthony Marino (Bristol-Myers Squibb), Daria Barwinska (Eli Lilly and Company), Aarti Shah (GlaxoSmithKline), and Natalya Alexander (Novartis Institutes for BioMedical Research, East Hanover, NJ).
Authorship Contributions
Participated in research design: Bentz, O’Connor, Bednarczyk, Coleman3, Lee, Palm, Pak, Perloff, Reyner, Balimane, X.Chu, Funk, Guo, Hanna, Herédi-Szabó, Hillgren, Li, Jamei, Lin, Neuhoff, Podila, Rajaraman, Salphati, Taub, Taur, Weitz, Wortelboer, Xia, Xiao, Yamagata, Zhang, Ellens.
Conducted experiments: Bednarczyk, Pak, Reyner, Balimane, Brännström, Guo, Hanna13, Herédi-Szabó, Li, Lin, Hollnack-Pusch, Mason, Patel, Podila, Plise, Rajaraman, Sands, Taur, Weitz, Wortelboer, Xia, Yabut, Yamagata.
Contributed new reagents or analytic tools: Bentz, O’Connor.
Performed data analysis: Bentz, O’Connor, Bednarczyk, Coleman, Lee, Palm, Pak, Perloff, Reyner, Balimane, Herédi-Szabó, Jamei, Taub, Ellens.
Wrote or contributed to the writing of the manuscript: Bentz, O’Connor, Bednarczyk, Coleman, Lee, Palm, Pak, Perloff, Reyner, Ellens.
Footnotes
- Received December 14, 2012.
- Accepted April 24, 2013.
↵1 Current affiliation: Novartis Institute for BioMedical Research, Cambridge, Massachusetts.
Work at Optivia was supported by Small Business Innovation Research Grants from the National Institutes of Health National Institute of General Medical Sciences [Grants R43GM086970-01, R43RR031474-01].
Disclaimer: The manuscript reflects the views of the authors and should not be construed to represent FDA’s views or policies. Lei Zhang has no conflict of interest to report.
↵This article has supplemental material available at dmd.aspetjournals.org.
Abbreviations
- A>B
- apical-to-basolateral
- AMP-PNP
- 59-adenylylimidodiphosphate
- B>A
- basolateral-to-apical
- Caco-2
- human colon adenocarcinoma cells
- DDI
- drug-drug interaction
- DMEM
- Dulbecco’s modified Eagle’s medium
- DMSO
- dimethyl sulfoxide
- LC-MS/MS
- liquid chromatography followed by tandem mass spectrometry
- LLC-PK1-MDR1
- Lilly Laboratories Cells - Porcine Kidney Nr. 1 cells transfected with MDR1 cDNA
- MDCKII-MDR1
- Madin-Darby canine kidney cells transfected with MDR1 cDNA
- Papp
- apparent permeability
- P-gp
- P-glycoprotein, also often referred to as MDR1 or ABCB1
- PCA
- principal component analysis
- TEER
- transepithelial electrical resistance
- U.S. Government work not protected by U.S. copyright