|Home | About | Journals | Submit | Contact Us | Français|
The purpose of the study was to assess the permeability of mouse blood-brain barrier (BBB) to a diverse set of compounds in the absence of P-glycoprotein (Pgp) mediated efflux, to predict it using an in combo PAMPA model, and to explore its role in brain penetration classification (BPC). The initial brain uptake (Kin) of 19 compounds in both wild-type and Pgp mutant [mdr1a(−/−)] CF-1 mice was determined by the in situ brain perfusion technique. PAMPA measurements were performed, and the values were used to develop an in combo model, including Abraham descriptors. Published rodent Kin values were used to enhance the dataset and validate the model. The model predicted 92% of the variance of the training set permeability. In all, 182 Kin values were considered in this study, spanning four log orders of magnitude and where Pgp decreased brain uptake by as much as 14-fold. The calculated permeability-surface area (PS) values along with literature reported brain tissue binding were used to group molecules in terms of their brain penetration classification. The in situ BBB permeability can be predicted by the in combo PAMPA model to a satisfactory degree, and can be used as a lower-cost, high throughput first-pass screening method for BBB passive permeability.
Discovery of central nervous system (CNS) drugs is challenging due to the complexity of neurological disorders, the paucity of pre-clinical models with proven human translational value, and the persistent difficulty of delivering drug molecules across the blood-brain barrier (BBB) to achieve optimal CNS exposure. There is an intricate debate taking place in the recent literature regarding how best to interpret the results obtained from in vitro, in situ, and in vivo blood-brain transport methodologies, to identify properties most relevant to successful CNS drug delivery, both in terms of the rate and the extent of penetration (Hammarlund-Udenaes et al. 2008; Martin 2004; Liu et al. 2005, 2008; Hitchcock 2008; Jeffrey et al. 2007; Cecchelli et al. 2007; Maurer et al. 2005; Pardridge 2007; Reichel 2009).
The brain-to-plasma distribution ratio has been widely used in drug discovery as an index of brain penetration (Kp = AUCtot,brain/AUCtot,plasma, where AUC is the area under the total drug concentration-time curve in the corresponding tissue over a dosing interval; also called log BB – see Glossary). However, Kp used in isolation is a misleading parameter since it is generally accepted that it is the unbound drug that exerts the physiological effect (Hammarlund-Udenaes et al. 2008; Martin 2004). Efforts to optimize CNS candidate selections based solely on Kp “may actually lead to unproductive/counterproductive design of drugs that are unnecessarily basic, lipophilic, and have greater degree of nonspecific partitioning into brain tissue” (Maurer et al. 2005). Increasing lipophilicity does not necessarily lead to increased efficacy (Summerfield et al. 2006), and may lead medicinal chemists astray in a chemical space that is hardly druggable due to poor solubility and metabolic instability.
A number of CNS drug discovery groups (Liu et al. 2005, 2008; Jeffrey et al. 2007; Summerfield et al. 2006) have proposed optimization strategies based on an integrated use of BBB permeability, the effect of efflux transport, and nonspecific plasma and brain tissue binding, to ultimately characterize brain penetration in terms of unbound drug brain distribution: Kp,uuin vitro = Cu,brain/Cu,plasma, where Cu,brain and Cu,plasma are the unbound concentrations in the brain (blood-free) homogenate and plasma, respectively. Ideally, a CNS drug would achieve a Kp,uuin vitro close to unity within a relatively short time to maximize CNS exposure. In practice, this would also maximize peripheral safety margins by minimizing differences in peak unbound concentrations between brain and plasma. For instance, BBB efflux makes Kp,uuin vitro < 1, requiring higher plasma concentrations to elicit a CNS effect.
In vivo microdialysis has been the long standing and labor-intensive standard methodology to measure unbound concentrations in living brain (Hammarlund-Udenaes et al. 2008; Fridén et al. 2007; Benveniste and Huttemeier 1990). A more refined consideration of the in vivo ratio, Kp,uu (cf., Glossary), draws a distinction between the concentration in the brain homogenate and the brain interstitial fluid (ISF), since the process of homogenization blends the intra- and extracellular brain fluids, which is relevant to compounds with poor cell membrane permeability that may distribute poorly to intracellular compartments. Nevertheless, there is cumulating evidence that suggests measurement of Kp,uuin vitro can be used to approximate free exposure in a pharmacologically meaningful manner. In vitro methods to assess the unbound drug fractions in plasma, fu,plasma, and brain tissue homogenates, fu,brain, have been extensively investigated (Liu et al. 2005, 2008; Jeffrey et al. 2007; Maurer et al. 2005; Kalvass and Maurer 2002; Mahar Doan et al. 2004; Kalvass et al. 2007; Wan et al. 2007; Summerfield et al. 2007).
Many different in vitro BBB models have been developed and proposed to measure permeability, some based on human brain endothelial cells (Terasaki et al. 2003; Garberg et al. 2005; Weksler et al. 2005; Abbott 2007). Despite limited successes, these models generally suffer from a significant loss in transporter function and poor tight junction organization (leakiness) as compared to the in vivo BBB. Many drugs companies simply use epithelial cells transfected with P-glycoprotein (MDCK or LLC-PK1) to conduct permeability screening in CNS drug discovery.
Kalvass et al. (2007) provided an integrated approach for assessing CNS steady-state distribution of drugs. They combined the Pgp efflux ratio, ER, with in vitro-measured unbound plasma and brain tissue concentrations. A plot of ER versus Cu,plasma/Cu,brain is divided into four brain penetration classifications (BPC) using 3-fold cutoffs as delineators.
In a similar example of integrative approach to better understand CNS exposure, Liu et al. (2005,2008) incorporated independent measures of BBB permeability by brain perfusion in their physiologically based pharmacokinetic (PBPK) model, and calculated brain equilibration half-time as t1/2,eq = const. / (PS·fu,brain), where the unbound brain fraction, fu,brain = Cu,brain/Ctot,brain. PS is the product of BBB permeability and the microcapillary specific surface area.
Given the number of compounds to test in early drug discovery, costly and labor-intensive in vivo measurements of CNS pharmacokinetic properties are impractical, and in vitro BBB models rely on relatively time-consuming cell culture. Useful in silico predictions of in vivo BBB transport properties could substitute for direct measurements in initial screening. In silico methods to predict Kp exist (Bendels et al. 2008, Zhang et al. 2008, Kortagere et al. 2008), but have not been applied to Kp,uu. The prediction of PSpassive has been the subject of a number of studies (Clark 2003; Abraham 2004; Lanevskij et al. 2008; Di et al. 2008). In silico attempts have been made to estimate the influence of Pgp on in vitro substrate transport (Didziapetris et al. 2003; Garg and Verma 2006).
The most reliable published PS data come from rodent in situ brain perfusion studies (Takasato et al.1984; Gratton et al.1997; Wu et al. 1998; Murakami et al. 2000; Bourasset et al. 2003; Cisternino et al. 2003, 2004; Youdim et al. 2004; Dagenais 2000; Dagenais at al. 2000, 2001, 2002, 2004, 2005; Cisternino et al. 2004; Liu, et al. 2005; Parepally et al. 2006; Bihorel et al. 2007; Obradovic et al. 2007). The modified Takasato method (Takasato et al. 1984) has been adapted to the mouse by Dagenais and Rousselle (Dagenais et al. 2000b), and applied to study BBB efflux using various knockout mouse models. This technique requires special surgical skills, is labor and animal intensive, and time consuming. Unless radiolabeled compounds are used, significant bioanalytical resources are required (e.g., liquid chromatography-mass spectrometry). Rodent brain perfusion studies are not performed routinely in the pharmaceutical industry.
As a cost-effective high-throughput alternative to in vitro and in situ methods to estimate BBB permeability, PAMPA (parallel artificial membrane permeability assay) could be used to assess PSpassive (Kansy et al. 1998; Avdeef 2003, 2005; Di et al. 2003,2008; Bermejo et al. 2004; Avdeef et al. 2005; Youdim et al. 2003; Ruell et al. 2004; Avdeef et al. 2004). PAMPA membranes are based on a mixture of phospholipids deposited into lipophilic microfilters, and contain net-negative lipid charge, mimicking some of the properties of endothelial cell membranes. Di et al. (2003) adapted porcine brain lipid extract dissolved in n-dodecane (2% w/v) as their PAMPA membrane barrier and demonstrated that molecules can be successfully binned into CNS+ and CNS− classes. The early attempts to establish a numerical relationship between PAMPA and Caco-2 cellular assays have shown considerable promise, with high linear correlations reported (Bermejo et al. 2004). An in combo PAMPA method (hybrid “combination” of in silico and in vitro methods) has been described, where it was possible to separate some active and passive components of cellular permeability (Avdeef et al. 2005).
In this study, we report 38 new in situ PS measurements performed on 19 compounds in wild-type (WT) and Pgp-deficient [mdr1a(−/−)] CF-1 mice (so called mutants equivalent in phenotype to the corresponding “knockout” model). One aim of the study was to quantify the influence of Pgp on PS values by comparing the mouse genotypes. The second aim was to use PSpassive values from the Pgp deficient data to train the in combo PAMPA procedure, followed by enhancement and validation based on additional published rodent values. We were interested in developing a practical, cost-effective, and rapid quantitative in combo assay which could be used in early screening for passive BBB permeability, and which could over time assist medicinal chemists with structure modification to improve a key determinant of CNS exposure for test compounds downstream in the discovery process, as part of an integrated BBB penetration model, incorporating PSpassive along with possibly ER, fu,brain, and fu,plasma. Finally we explored the implicit contribution of passive BBB permeability in the brain penetration classification (BPC), which has not been discussed in the published literature.
The unlabeled chemicals used in this study (training set) were purchased from Sigma-Aldrich/RBI (St. Louis, MO, USA) and Tocris Bioscience (Ellisville, MO, USA), and used as received, except that analytical-grade amitriptyline and indinavir (Merck), astemizole, domperidone and galanthamine (Janssen-Ortho), clozapine (Novartis), gabapentin and sertraline (Pfizer), ritonavir (Abbott), saquinavir (Roche), and terfenadine (Aventis), were kindly provided by their manufacturer. PAMPA lipid (lecithin mixture) was obtained from pION (PN 110669), and was stored at −20°C when not used. The pH of the assayed donor solutions was adjusted with a universal buffer (pION, PN 100621). A buffer solution at pH 7.4 containing a chemical scavenger to simulate tissue binding and maintain sink conditions (pION ASB-7.4 buffer, PN 110139) was used as the receiver solution.
Adult male CF-1 mice (wild type and mdr1a(−/−), 30-40 g, 6-8 weeks old) were obtained from Charles River Laboratories (Wilmington, MA, USA). Animals were housed in a room with a controlled environment (22 ± 3° C; 55 ± 10% relative humidity) and maintained under a 12-h dark:light cycle (light from 6 a.m. to 6 p.m.). Animals had access to food and tap water ad libitum. All animal experiments were evaluated and approved by the Institutional Committee for Good Animal Practices (UQAM, Montréal, Québec, Canada).
The in situ mouse brain perfusion has been described in detail elsewhere (Dagenais 2000; Dagenais et al. 2000). Briefly, mice were anesthetized with intraperioneal ketamine/xylazine (140/8 mg/kg). The right hemisphere was perfused through the right common and internal carotid arteries following ligation of the external branch. The cardiac ventricles were severed immediately before brain perfusion (Krebs-bicarbonate buffer gassed to pH 7.4 with 95% O2 and 5% CO2 at 37° C) at 2.5 mL/min via a syringe pump. Test compounds were spiked in perfusate at a targeted concentration of 1 μM from a DMSO or dilute acetic acid stock solution (400-fold dilution). The perfusions were terminated by decapitation. Single time point experiments were performed (1, 2 or 3 minutes) based on the molecular properties of the compounds and their expected permeability category (low, intermediate, high). Previous studies were used to gauge the length of perfusion most likely to be within the linear portion of the uptake curve. The brain was removed from the skull and dissected on ice to isolate the right hemisphere. At this point, a sample of perfusion fluid was collected at the tip of the catheter by activation of the infusion pump. Samples were frozen in liquid nitrogen and maintained at −80°C until analysis by LC/MS or LC/MS/MS.
The generic sample preparation and bioanalysis procedures have been reported in detail elsewhere (Dagenais et al. 2002, 2004). As appropriate, samples of perfusion fluid were acidified with 1% v/v glacial acetic acid, and diluted with the mobile phase used at the start of gradient elution without further preparation. Brain samples (ca. 150 mg) were homogenized in 4 volumes of aqueous buffer, followed by protein precipitation with 2 volumes of ice-cold acetonitrile. Following centrifugation, the supernatant was either injected directly onto the HPLC system, or evaporated and reconstituted in 40% acetonitrile/ 60% water with 0.1% formic acid. All in vivo samples were quantified by LC/MS or LC/MS/MS using an Agilent 1100 LC/MSD system (Agilent Technologies, Wilmington, DE), or a PE-Sciex API-3000 MS/MS system (Applied Biosystems, Foster City, CA) fitted with a Shimadzu LC-10 HPLC system (Shimadzu Scientific Systems, Columbia, MD) and a CTC-Pal autosampler (Leap Technologies, Carrboro, NC). Gradient elution (1 to 2 mL/min) using admixtures of 0.1% formic acid or 25 mM acetate, and acetonitrile or methanol , was applied to high performance analytical columns (Agilent Zorbax C-18, YMC NH2 or Agilent Zorbax Extend C-18, all 4.6 mm × 50 mm). Parent ion monitoring (i.e., selective ion monitoring, MS) or multiple reaction monitoring (MS/MS) were used for mass spectrometric detection. Diclofenac was used as the internal standard. Typical run times were between 2 and 8 minutes per sample. Calibration curves were fit using linear or quadratic relationships with standard software from the instruments' manufacturer. A deviation of 20% or less from the predicted curve was considered acceptable for points on the calibration curve. Samples were re-assayed after appropriate dilution if the concentration could not accurately be interpolated or extrapolated.
Initial uptake data from brain perfusion experiments can be assimilated to a pharmacokinetic tissue uptake clearance (Dagenais 2000; Dagenais et al. 2000), but traditionally have been referred to as Kin values, which can be estimated from single time point experiments using the following relationship (Dagenais 2000; Dagenais et al. 2000), Kin = ( Xbr / T ) / Cpf, where Xbr is parenchymal brain concentration (mol/kg units) of the tracer, Cpf is the perfusion fluid (analogous to arterial) concentration (mol/L) of the test compound collected at the tip of the infusion catheter, T is the perfusion time (min), and Kin is the initial rate of brain uptake. Parenchymal brain concentration is obtained by subtracting from the total concentration, the contribution from the vascular volume, Xbr = Xtotal − Vvasc Cpf. A vascular volume of 1 mL/100g was used to correct total brain concentrations.
Using the Crone-Renkin equation, the rate constant is defined as Kin = Fpf ( 1 − e −PS/Fpf ), where Fpf is the regional cerebral flow of perfusion fluid, and PS is the capillary permeability-surface area product. Usually, diazepam is used as a flow marker to determine Fpf, since its uptake clearance is expected to be maximal (Dagenais 2000; Dagenais et al. 2000). However, in the training set used in this study, three compounds (amitriptyline, buspirone, chlorpromazine) had brain uptake clearances that were greater than that of diazepam (255 mL/100 g/min in the mouse; Dagenais 2000; Dagenais et al. 2000). Since the results tend to be more variable and sensitive to changes in experimental conditions under a flow-limited situation, the regional cerebral flow, Fpf, was taken to be 430 mL/100g/min for the purpose of the training set conversions. This value was selected so that all of the measured Kin may be converted to PS values. This choice did not alter the ranking of the PS values, as considered below.
In order to apply the Abraham descriptors, the PS value were temporarily converted to the intrinsic scale (Avdeef et al. 2005), according to the equation PoBBB = (PS / S) ( 1 + 10±(pH-pKa)), with the '+' used for acids, and the '−' sign used for bases, and where S = brain capillary surface area, assumed to be 100 cm2/g (Ohno et al. 1978).
In addition to Pgp deficient (mdr1a(−/−)) data, wild-type mouse uptake clearance data were measured in this study. The results from Pgp-deficient mice were expected to result in a better estimate of passive permeability. The Pgp effect is defined by the ratio of the Pgp-deficient uptake clearance to that of wild-type up uptake clearance.
The PAMPA Evo instrument from pION INC (Woburn, MA, USA) was used in this study. The 96-well microtitre plate “sandwich” (pION, PN 110212, preloaded with magnetic stirrers) filters were coated with a 20% (w/v) alkane solution of a lecithin mixture containing an excess of negatively-charged constituents (pION, PN 110669). Sample concentrations in the buffer solutions for the compounds with low-UV absorption were about 500 μM (e.g., gabapentin, DPDPE, meperidine). However, for most of the compounds, UV sensitivity was good, and the typical concentrations were about 50 μM. DMSO-free stock solutions were prepared for some of the compounds (buspirone, colchicine, quinidine, tolbutamide, U69593, fentanyl, sertraline, ritonavir, astemizole, clozapine, deltorphin II, DPDPE, gabapentin, galanthamine, indinavir), to improve on UV sensitivity in the 210-240 nm part of the spectrum. The donor solutions were varied in pH (NaOH-treated universal buffer), while the receiver solutions had the same pH 7.4. This was necessary in order to correct the effective permeability values for ionization and aqueous boundary layer (ABL) effects (Avdeef et al. 2004, 2005). The pKaflux-centered (see below) gradient pH values were selected by the pOD procedure (Ruell et al. 2003). The receiver solutions contained a surfactant mixture (“lipophilic sink”) to mimic tissue binding (Avdeef 2003). Vigorous stirring was employed in the assay, with stirring speed set to produce an ABL thickness < 50 μm, to minimize the ABL contribution to the measured permeability. The PAMPA sandwich was assembled and allowed to incubate for 30 minutes for highly permeable molecules (e.g., amitriptyline, chlorpromazine, loperamide, sertraline, probenecid and verapamil), and 15 h for poorly permeable molecules (e.g., galanthamine, DPDPE, deltorphin II, indinavir), in a controlled-environment chamber (pION PN 110205) with a built-in magnetic stirring mechanism. Both the donor and receiver wells were assayed for the amount of material present, by comparison with the UV spectrum (210 to 500 nm) obtained from a reference standard. Permeability values were corrected for membrane retention (Avdeef 2003).
In brain endothelial microcapillaries, the ABL thickness is likely to be < 1 μm (given that the radius of the human microcapillaries is about 3 μm), whereas in unstirred PAMPA, the ABL thickness can be as high as 4000 μm (Avdeef et al. 2004). By taking PAMPA (stirred or unstirred) data over a range of pH, it is possible to correct for the effect of the ABL, by applying the pKaflux method (Avdeef 2003; Avdeef et al. 2004, 2005; Ruell et al. 2003), briefly described below.
In the analysis of PAMPA data, four types of permeability values are considered: (a) the effective permeability coefficient, Pe (which is directly determined); (b) the aqueous boundary layer permeability, PABL, based on the resistance of the water layer adjacent the membrane; (c) the membrane permeability, Pm (which depends on pH with ionizable molecules); and (d) the intrinsic permeability, Po, which is the permeability of the uncharged species (the maximum possible value of Pm at the pH where an ionizable molecule is uncharged). ABL-limited transport is often observed for lipophilic molecules, when Po >> PABL. The basic equation describing permeability is (Avdeef et al. 2005)
which is valid provided that Po > 10 PABL (lipophilic molecules). The maximum possible effective (measured) permeability, Pemax, is defined as log Pemax = log PABL − log ( 1 + PABL / Po ). When Po >> PABL (highly permeable molecules), Pemax, ≈ PABL. The “flux” ionization constant, pKaflux, refers to the pH value where the resistance to transport across a permeation barrier is 50% due to the ABL and 50% due to the membrane (Avdeef 2003).
Our core computational object was to model PSpassive values. From a survey of the published literature, 507 PS values were identified, based on in vivo intravenous injection (i.v.), bolus carotid artery injection (BUI), and in situ brain perfusion methods, for rats, mice, guinea pigs, rabbits, dogs, and cats. It was decided to focus only on rat (n=335, 66%) and mouse (n=131, 26%) data, accounting for about 92% of the collected values.
It is hypothesized here that mouse and rat data are comparable for the purposes of our prediction study. This may be reasonable, as suggested by Murakami et al. (2000). For their reported rat and mouse permeability values (21 compounds), the linear regression of rat vs. mouse logarithm of the intrinsic (defined below) permeability coefficients yielded the intercept = 0.21 ± 0.24 and slope = 1.01 ± 0.05.
Since plasma protein binding lowers values of PS (in comparison to protein-free perfusate experiments), i.v. data were not used for lipophilic compounds. Compounds that had reported saturable transport were also excluded. Since we were interested to select in situ data as free of efflux effects as practical, we chose our training set PS values from studies which used some sort of transport inhibition (e.g., Pgp-KO, cyclosporine A, PSC833, GF120918, self-inhibition by using high concentrations). Also, simple amino acids and dipeptides were excluded, except for those with reported non-saturable Kd values. In our own data, the L-isomer of p-F-phenylalanine was excluded from the training set, since comparison with the D-isomer clearly indicated that its transport was facilitated. As an added pruning, several compounds were excluded from the training set on the suspicion that carrier-mediated transport was possible, by comparisons of structures to known transport-mediated molecules. Out of the starting set of 507 PS values, a total of 182 “passive” rodent values were selected in our study, including the in situ measurements reported here.
The in combo PAMPA model was trained with a set of 80 compounds: 31 bases, 11 acids, 15 ampholytes, and 23 neutral molecules (encompassing 130 different measured PS values). In addition to the 130-PS training set, an additional 52 molecules (Summerfield et al. 2007; Gratton et al. 1997; Obradovic et al. 2007) were taken to serve as the test set. Table 1 contains physical properties and the Abraham (2004) descriptors of the training and test set molecules.
Abraham's linear free energy relations (LFER) applied to a BBB permeability model may be stated as (Abraham 2004)
where c0…c5 are the multiple linear regression (MLR) coefficients, and where R (dm3 mol−1 / 10; also called E) is the excess molar refraction, which models dispersion force interaction arising from pi- and n-electrons of the solute, π (also called S) is the solute polarity/polarizability due to solute-solvent interactions between bond dipoles and induced dipoles, α (also called A) is the solute H-bond acidity, β (also called B) is the solute H-bond basicity, and Vx is the McGowan molar volume (dm3 mol−1 / 100) of the solute.
Eq. (2) uses intrinsic BBB permeability values, PoBBB, rather than PS values. This is because the Abraham molecular descriptors have been developed for uncharged species, and so it was decided to convert all permeability values (PS and PAMPA) to intrinsic values, PoBBB and PoPAMPA, in order to develop the computational model. This may seem unnecessary, given that the environment of the BBB is very close to pH 7.4. However, the transformation is solely a computational strategy, in order to take full advantage of the Abraham descriptors. After the model was developed, the calculated intrinsic values were converted back to PS scale. In effect, by these transformations, we have adapted the Abraham molecular descriptors for charged molecules.
In addition to the LFER model, we explored how well PAMPA measurements, augmented with one (or two) of Abraham's solvation molecular descriptors, can predict PSpassive values. The combination of measured PAMPA and a calculated LFER descriptor defines the in combo method below. A trial-and-error testing of various combinations of one or two Abraham descriptors and PAMPA was tried, using
where a0, a1… are MLR coefficients. The usefulness of such an approach has been discussed elsewhere (Avdeef et al. 2005). Fewer MLR coefficients are necessary in eq. 3, compared to eq. 2, because PoPAMPA already encodes for some of the transport characteristics common to the permeability models.
The best prediction model was validated by testing its ability to predict BBB permeability of compounds not used in the training set.
The Abraham descriptor calculation and the computational model testing used the Algorithm Builder V1.8 and ADME Boxes V4.9 computer programs (Lanevskij et al. 2008; Didziapertris et al. 2003) from Pharma Algorithms (Toronto, Canada). Since the 52-molecule test set and ten of the zwitterionic training set molecules were not available to us for PAMPA measurement, the pCEL-X program (pION) was used to predict the missing PAMPA values.
The PS parameter is an indicator of early access into the CNS. The short time over which the measurement is performed minimizes egress from brain to blood. As such, this parameter is independent from, and cannot predict the extent of brain distribution at steady-state, even if there is some overlap in the properties that impart high affinity for brain tissue and high passive BBB permeability. Two additional considerations further support this statement. First, the upper limit of the observed BBB permeability in vivo is cerebral blood flow. Therefore, compounds may have a similarly high permeability, but very different affinity for brain tissue, as illustrated by the lipophilic amine sertraline which has a very high brain volume of distribution, in comparison to the neutral compound diazepam (low to intermediate). Second, apparent permeability is the net result of passive and when applicable, carrier-mediated transport components, as indicated in Table 2. Hence a Pgp substrate may have low apparent permeability, but high affinity for brain tissue, which would obscure any underlying correlation that may have existed with respect to passive permeability.
The Pgp effect (Table 2) is an index of the magnitude of efflux under initial uptake conditions. It has been suggested than under initial brain uptake conditions and in the absence of plasma protein binding, the Pgp effect is an underestimate of the steady-state situation (Dagenais et al. 2001). Moreover, deriving the Pgp effect from the nonlinear PS value scale (in contrast to Kin) would yield much higher ratios for compounds that go from low or intermediate extraction to high extraction when efflux is removed, so it should be used to rank compounds based on their efflux potential, not in absolute terms. The compounds represented a wide range of Pgp effects ranging from no interaction (ratio of 1) to a 14-fold difference (based on Kin values), with Kin values spanning four log orders of magnitude.
Table 1 lists the PAMPA intrinsic permeability values. The values for doxorubicin and gabapentin are poorly determined, as indicated by the standard deviations (SD). This was due to the low UV absorbance of the molecules. Figure 1 shows the PAMPA log Pe as a function of pH for twelve of the studied molecules. Morphine, naltrindole, and bremazocine are ampholytes, and are characterized by parabolic-shaped profiles. The bases in Figure 1 are characterized by an ascending hyperbolic curves with increasing pH.
The best-fits of eq. 1 to the data are represented by the solid curves, and the derived lipid membrane-based log Pm - pH curves are represented by dashed curves. The horizontal dotted lines correspond to the log PABL values, resulting from the regression analysis based on eq. 1. The maximum point in the log Pm curves corresponds to the intrinsic permeability coefficient, log Po.
The PAMPA method described here may appear to be low-to-medium throughput, since for most of the compounds, permeability was determined in 3-12 different pH buffers (cf., Fig. 1). This was done to reconcile the substantial difference between the aqueous boundary layer thickness in PAMPA (1500-4000 μm in unstirred plates) and in the in situ assay (< 1 μm). Some pharmaceutical companies perform PAMPA measurements in duplicate at one pH in high-throughput assays (often without stirring). To match such a throughput, it can be proposed here that PAMPA be done at two pH values as singletons with the pH values selected to straddle the pKaflux value, as described by Ruell et al. 2003. This would be sufficient to correct the data for the effects of the aqueous boundary layer, to better match the in situ conditions. Furthermore, if vigorous stirring were used for lipophilic compounds during the assay, the method could have up to 100-fold increase in the dynamic range of the effective permeability, with a concomitant decrease in the permeation time. Such a proposed procedure would have nearly the same workload as the commonly used high-throughput protocol, and thus could be considered high throughput. (Given that PAMPA values themselves can be predicted (e.g., pCEL-X), our proposed PSpassive prediction model can be done entirely as a very fast in silico method, perhaps suitable for ranking molecules in virtual compound libraries.)
Comparisons of log PoBBB and log PoPAMPA were explored according to eq. 3, using various combinations of “booster” LFER descriptors. Combinations of log PoPAMPA and the H-bond Abraham descriptors produced the lowest r2 values. Using both α and β descriptors, the best-model multiple linear regression coefficients resulting from the search produced r2 = 0.80 and s = 0.79 (Table 5). By comparison, the five-descriptor Abraham model yielded r2 = 0.69 and s = 0.99 (Table 5). Inspection of the residuals suggested that the acid compounds appeared to have higher than expected passive permeability, compared to the other compounds in the training set. This may be due to the anisotropic properties of the biological membranes, consisting of various bilayer-forming amphiphilic lipids and membrane-anchored proteins. The distribution of the lipids and proteins is complex and uneven in the BBB membranes. It may be that acids, bases, neutral molecules, and zwitterions undergo passive diffusion that can best be described by different mixes of physical property descriptors. To test this hypothesis, the 130-PS training set was then partitioned into four groups: acids, bases, neutrals, and zwitterions.
The partitioned-set regression analyses led to an improved overall model, and the individual set parameters are listed in Table 5. The anions and cations had the highest correlation coefficients, 0.90 and 0.86, respectively. For the zwitterions, the 0.21 coefficient for the log PoPAMPA term was considerably lower than those of the other three groupings. Also, the contribution due to H-bond donor strength (α) varied extensively across the four types of charge groups, with neutral compounds appearing to have enhanced BBB permeability with increased H-bond donor strength, contrary to expectations. The zwitterion set had the poorest r2 of the four groups. Abraham et al. (1997) noted that the LFER descriptors were not ideally suited for these charged species. Also, the log PoPAMPA descriptors for many of the zwitterions were calculated using pCEL-X (whereas those of the other training-set molecules were measured), and thus are less accurate. After the in combo model was developed, the calculated results were transformed back to the PS scale.
Figure 2a is a plot of observed and calculated intrinsic permeability coefficients, using the charge-partitioned models in Table 5. The combined statistics indicate r2 = 0.92 and s = 0.51. Figure 2b is the plot of the observed log PS as a function of the calculated values (filled circles). The transformed PS values indicated r2 = 0.80 and s = 0.51.
Figure 2 (yellow squares) shows the relationship between the in combo model predicted and observed test set permeability values (Summerfield et al. 2007; Gratton et al. 1997; Obradovic et al. 2007) from Table 4. For the charge-partitioned prediction, r2 = 0.59, and s = 0.67 in the PS form (Fig. 2b), and r2 = 0.82 in the intrinsic permeability form (Fig. 2a). This may be considered an adequate validation of the in combo training set. The largest test-set residuals were seen with AZ13007, pemoline, and lamotrigine.
Figure 3 shows a rank-ordered comparison of predicted and observed in situ PS values, using a “low,” “intermediate,” and “high” designations. The classification boundaries were selected by inspection, and in the figure, the PS range (in 10−4 mL/g/s units) 0 to 20 defines the low class; the high class is defined by PS above 70 for the predicted set and above 150 for the observed set. The intermediate class is the zone between the high and low boundaries.
For the training set (Table 3), there were three false negatives (galanthamine, buspirone, and terfenadine) and four false positives (vinblastine, gabapentin, naltrindole, and ritonavir). This corresponds to about 18% misclassification. The origin of these discrepancies is unclear. The rest of the training set compounds (82%) were correctly classified: 37% as high, 11% as intermediate, and 34% as low.
Also included in the figure are some of the test-set molecules, indicated by italic text. These were similarly distributed as those of the training set.
The underlined compounds in Figure 3 were those of the training set, where both the wild-type (not used in model training) and the Pgp-deficient results were predicted in the same class. Largely, these are compounds that do not appear to be substrates of Pgp in the in situ method.
The seven compounds designated with an asterisk in Figure 3 are based on wild-type PS values (not used in the training), and reveal a down-step in classification. The corresponding values of the Pgp effect ratios (Table 2) are: quinidine 14.2, ritonavir 8.3, loperamide 10.4, verapamil 6.4, diltiazem 2.0, SNC121 8.6, and methadone 2.6.
Despite the fact that some training set compounds are known substrates of putative transport systems in the BBB, the model performs relatively well. Even if there is a non-Pgp efflux, or active or facilitated uptake process, the molecular properties of the compounds (and apparent permeability values) span a much broader range than the efflux ratios at the BBB. For this reason DPDPE, probenecid, gabapentin, tolbutamide, and M6G were not excluded, even though these are known to be substrates of transport systems (Deguchi et al.1997; Sun et al. 2001; Takanaga et al. 1998; Luer et al. 1999; Uchino et al. 2002; Su et al. 1995; Ohtsuki and Terasaki 2007). DPDPE was shown to be substrate of a saturable uptake system (Dagenais et al. 2002). Passive diffusion of DPDPE is negligible without this transport system. The estimated transport parameters, along with the highest concentration tested, 150 μM, suggest a Kin of 0.0862 mL/100g/min for DPDPE, which may be reasonable to use, in place of the higher value in Table 2. Probenecid is an organic anion efflux substrate and inhibitor (Deguchi et al. 1997; Sun et al. 2001). Since gabapentin is a substrate of the large neutral amino acid transporter (LNAAT; system L), the Kin value in Table 2 probably overestimates passive permeability. This is nicely illustrated by the difference in BBB permeability between p-F-L- and p-F-D-phenylalanine in Table 2; due to its enantiomeric form, only the former is a recognized substrate of the LNAAT. Tolbutamide appears to be a non-Pgp efflux substrate (Takanaga et al. 1998). M6G appears to have an uptake transporter (Bourasset et al. 2003). Quantitatively, how transporters affect the blood-brain transport of these compounds in relation to intrinsic BBB permeability should be further explored as more in vivo data become available.
It's clear that there is a cost and speed advantage in using in combo PAMPA to predict BBB permeability, compared to in vitro, in situ or in vivo methods. Since PAMPA can only address the passive permeability component in the overall transport process, it's best to use it for virtual screening of large sets of compounds, or in a supportive role. For example, when an in vivo, in situ, or in vitro permeability measurement is made, it may be difficult to be certain to what extent the result represents a passive, a carrier-mediated, or an active influx/efflux process. Combining the PAMPA-predicted permeability with the more biomimetic measured values can sometimes shed light on the contributions of several components of transport, as illustrated previously (Bermejo et al. 2004; Avdeef et al. 2005). The two examples below illustrate the use of PSpassive determined by the in combo PAMPA, in the first case to indicate transporter effects (influx and efflux), and in the second case to indicate BPC classification using different parameters from those used by Kavlass et al. (2007)
Compounds that are Pgp specific (Pgp effect > 1.5 in Table 2) will have attenuated brain uptake. Their predicted PSpassive will exceed the measured PS value, as illustrated by the asterisked compounds in the ranking scheme in Figure 3. Figure 4 shows a quantitative plot of the Pgp effect with the down-pointing (blue) triangle symbols. For this set of compounds, the vertical axis represents the measured PS values (WT), whereas the horizontal scale represents the PSpassive values calculated by the in combo PAMPA model, drawing on training set of measurements performed with KO mice or rodents where some inhibition of transport was implemented. The further the down-triangle symbols are displaced from the solid diagonal line, the more the compound indicates a Pgp effect. The dashed diagonal lines represent a threefold difference, where compounds falling below the threshold line show a significant Pgp effect.
Compounds that are actively uptaken in in situ brain perfusion measurements are indicated as upward displacements by upward-pointing (red) triangles in Figure 4. The calculated PSpassive underestimates the observed permeability values. Most of the examples are amino acids. But there are some other examples, e.g., doxorubicin, daunomycin, and mitoxanthrone. These compounds have been extensively studied in terms of efflux specificity, but very little has been reported regarding their being influx transported by a facilitated process.
The BPC scheme described by Kalvass et al. (2007) assesses the role of Pgp and other putative processes in “impairing” (or compensating) brain penetration. Their scheme was presented in terms of a plot of the Pgp efflux ratio (ER), defined as the ratio of Kp between KO and WT mice, versus the ratio of the unbound drug concentrations in plasma and brain tissue (the inverse of Kp,uuin vitro). The horizontal and vertical lines indicate an efflux ratio of 3, and ratio of unbound plasma-to-brain concentrations of 3, respectively. The different quadrants delineate various CNS penetration characteristics: 1) impaired by Pgp and/or other active process(es); 2) impaired by non-P-gp mechanism; 3) no impairment; and 4) Pgp, but no impairment due to compensatory mechanism.
Their BPC plot for 34 compounds is reproduced in Figure 5, but with two additional parameters, PSpassive (determined by in combo PAMPA) and t1/2,eq (calculated from PSpassive and fu,brain using the method and data from Liu et al. 2005) superposed on the compound symbols: (a) the size of the symbols is proportional to log PSpassive; (b) the values of log t1/2,eq are proportionally represented by the “cross” symbol, with the largest cross symbol representing 624 min (ivermectin) and the smallest cross symbol representing 2 min (meperidine). We acknowledge that PSpassive is an overestimate of the true in vivo permeability for efflux substrates, which at first consideration appears to bias the estimate of t1/2,eq by shortening it. However, while a decrease in BBB permeability due to Pgp efflux may be expected intuitively to result in lengthening of t1/2,eq modeling of blood-brain transport in this specific case suggests that the effect would be more or less reversed depending on the magnitude of the increase in blood-to-brain egress (Syvänen et al. 2006). Although the relative balance between these two parameters is unknown for the Pgp substrates in Figure 5, we still considered the exercise conceptually useful, especially to qualify Class 3 compounds.
In the BPC plot in Figure 5, Classes 3 and 4 on the left side are characterized by unbound drug concentration in the brain of about 30% or more that in the blood. These two classes may be designated “CNS+”. Conversely, the other classes on the right side of the plot can be designated “CNS−“ for the purpose of the present discussion. Di et al. (2003) were able to differentiate CNS+ from CNS− groups using PAMPA. In the present study, PSpassive was used to define class boundaries in the BPC scheme in Figure 5.
Inspection of Figure 5 indicates that the most intrinsically permeable compounds are associated with BPC Class 3 (no impairment to brain penetration). However, intermediate to high PSpassive values also are associated with Class 1 (Pgp impaired penetration), simply reflecting that PSpassive is not a predictor of Pgp efflux ratio. However, it should be noted that for compounds that can bind to Pgp, the residence time in the cell membrane (a direct function of PSpassive) must be sufficient for an interaction to occur and result in significant efflux. For instance, although some high permeability compounds may display affinity for Pgp and act as inhibitors, they display little or no apparent efflux due to passive escape. The three-fold off-diagonal (outside of the dashed-line boundary in Fig. 5) compounds generally have very low PSpassive, suggesting it is also a contributor to poor CNS penetration. Evidently PSpassive alone cannot differentiate impaired from non-impaired compounds in the BPC scheme. The average values of PSpassive, fu,brain and t1/2,eq (a combination of the first 2 parameters) for the 34 compounds in Figure 5 were compiled for each of the BPC groups, along with the 68% probability ranges, and are listed in Table 6. Conceptually, t1/2,eq is shortest when both permeability and free fraction in brain are high. As mentioned above, there is some overlap in the molecular properties that govern the latter two parameters, but that can be obscured by active efflux or uptake. There cannot be permeability if the compound is unable to partition in cell membranes, which is to a large extent analogous to partitioning in brain tissue. The optimal CNS candidate has good permeability, moderate affinity for brain tissue and no efflux, which appears to be the hallmark of Class 3 compounds. One notes that Class 1 (impaired) and Class 3 (unimpaired) have large differences in the average PSpassive (68 vs. 430, 10−4 mL/g/s units), perhaps suggesting that these compounds have a slightly longer residence time in the luminal cell membrane that maximizes interactions with Pgp as discussed above. From the PS parameter ranges, 170 appears to be a good boundary value to differentiate the two classes. Another important difference between the two classes is the value of brain tissue binding. This binding in Class 1 compounds is notably greater (average fu,brain = 6%) than those in Class 3 (average fu,brain = 19%), but the two ranges overlap significantly. Perhaps these two measured quantities could be used in drug discovery in the following way: a compound is very likely not to be brain penetration impaired (Class 3) if PSpassive > 170 × 10−4 mL/g/s and fu,brain > 16%. Class 4 (unimpaired due to compensation mechanisms) and Class 3 (unimpaired) both have high PSpassive, with considerable overlap in permeability values. However, the average values of fu,brain for the two classes are quite different. In such a comparison, a test compound is more likely to be Class 4 than Class 3 if fu,brain < 3.4% and PSpassive < 180 × 10−4 mL/g/s. Classes 2 and 1b are characterized by low values of PSpassive, again suggesting that this parameter contributes at least in part to poor brain penetration. However, the values of fu,brain are dramatically different in the two classes (Table 5).
Although the overall number of compounds used to create the BPC probably has to be increased in order to confirm the rules we suggest, the empirical observations above may be useful in drug discovery applications, since PSpassive (PAMPA) and fu,brain (brain homogenate dialysis) can be measured at high-throughput speeds at a relatively low cost.
There exist multiple and complementary methodologies to study BBB transport and distribution in various brain compartments. The ability to cross the BBB at a sufficiently high rate (i.e., permeability) to avoid excessive blood-brain equilibration delays, unhindered by efflux processes, is a pre-requisite for CNS active compounds to distribute to brain, and exert their pharmacological action while maximizing peripheral safety margins. We used a mouse brain perfusion technique to estimate the BBB permeability of drugs and drug-like compounds with a wide range of molecular properties, and enhanced this dataset with published rodent values. We have demonstrated that the in combo PAMPA method, trained with 130 in situ P-glycoprotein deficient/inhibited rodent brain perfusion data, was able to predict 82% of the variance in the intrinsic BBB permeability of 52 external test compounds, whose permeability values were not used in the original training. Our investigation, based on 182 rodent brain perfusion results, is one of the largest PS-based published study to date used to develop a BBB permeability prediction model. The passive permeability predictions allowed us to further qualify and understand the brain penetration classification suggested by other investigators. Based on correlation plots (Fig. 2), the rank order comparisons (Fig. 3) and the transporter effect considerations (Fig. 4), we conclude that we have met the objectives of our study: to develop a practical, low-cost, and fast quantitative method which could be used for early passive BBB permeability screening, and for assisting medicinal chemists with structure modification to improve the BBB permeability of test compounds downstream in the CNS drug discovery process.
We thank Debra McCombs, Julie Zalikowski and Cindy Shen for their skilled bioanalytical support at AstraZeneca. Discussions with, and suggestions from, Drs. Joan Abbott, David Begley, and Sarah Thomas of King's College London, Khanh Bui and Mary Bock from AstraZeneca, and Michel Demeule at UQAM, Per Nielson from pION, are gratefully acknowledged. Part of this work was supported by Grant Number R44MH75211 from the National Institutes of Health (to pION). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institute of Mental Health or the National Institutes of Health.
Contribution number 18 in the PAMPA - a Drug Absorption in vitro Model series from pION. Ref. 47 is part 17 in the series.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.