Prediction of Drug Permeability Using In Vitro Blood–Brain Barrier Models with Human Induced Pluripotent Stem Cell-Derived Brain Microvascular Endothelial Cells

Abstract The strong barrier function of the blood–brain barrier (BBB) protects the central nervous system (CNS) from xenobiotic substances, while the expression of selective transporters controls the transportation of nutrients between the blood and brain. As a result, the delivery of drugs to the CNS and prediction of the ability of specific drugs to penetrate the BBB can be difficult. Although in vivo pharmacokinetic analysis using rodents is a commonly used method for predicting human BBB permeability, novel in vitro BBB models, such as Transwell models, have been developed recently. Induced pluripotent stem cells (iPSCs) have the potential to differentiate into various types of cells, and protocols for the differentiation of iPSCs to generate brain microvascular endothelial cells (BMECs) have been reported. The use of iPSCs makes it easy to scale-up iPSC-derived BMECs (iBMECs) and enables production of BBB disease models by using iPSCs from multiple donors with disease, which are advantageous properties compared with models that utilize primary BMECs (pBMECs). There has been little research on the value of iBMECs for predicting BBB permeability. This study focused on the similarity of iBMECs to pBMECs and investigated the ability of iPSC-BBB models (monoculture and coculture) to predict in vivo human BBB permeability using iBMECs. iBMECs express BMEC markers (e.g., VE-cadherin and claudin-5) and influx/efflux transporters (e.g., Glut-1, SLC7A5, CD220, P-gp, ABCG2, and MRP-1) and exhibit high barrier function (transendothelial electrical resistance, >1000 Ω × cm2) as well as similar transporter expression profiles to pBMECs. We determined that the efflux activity using P-glycoprotein (P-gp) transporter is not sufficient in iBMECs, while in drug permeability tests, iPSC-derived BBB models showed a higher correlation with in vivo human BBB permeability compared with a rat BBB model and the Caco-2 model. In a comparison between monoculture and coculture models, the coculture BBB model showed higher efflux activity for compounds with low CNS permeability (e.g., verapamil and thioridazine). In conclusion, iPSC-BBB models make it possible to predict BBB permeability, and employing coculturing can improve iPSC-BBB function.


Introduction
Clinical trials of central nervous system (CNS) drugs have shown a very low overall success rate (6.2% vs. 13.3% for non-CNS drugs) and the time required for approval by the U.S. Food and Drug Administration in resolving these issues is the development of in vitro assays, as an alternative to animal models, which can accurately estimate pharmacokinetics in the CNS. 3 In silico simulation models have been established to predict in vivo blood-brain barrier (BBB) permeability. The accuracy of in silico prediction of BBB permeability to small molecules (<1000 Da) has improved; however, it remains difficult to predict BBB permeability to noncovalent, inorganic, higher molecular weight, and mixtures of compounds using this model. 4 The BBB is a key structure in the CNS for nutrients and drugs to penetrate from the blood vessels to the brain cortex, and the tight junctions (Tjs) formed by brain microvascular endothelial cells (BMECs) have a high barrier function (*1800 O · cm 2 in vivo in TEER: transendothelial electrical resistance) that protects the brain from neurotoxic substances. 5 Conversely, the high barrier function of BMECs hinders the ability of neurotherapeutic drugs to exert therapeutic effects in the CNS. To investigate CNS pharmacokinetics, although animal models are frequently used, transporter expression in the BBB differs among animal species, 2 which affects drugs penetration through the BBB. To date, in addition to animal models, Transwell permeability assays, such as the parallel artificial membrane permeability assay (PAMPA) and Caco-2 permeability assay, are used to predict in vivo BBB permeability; however, the predictive accuracy of these assays is insufficient due to use of non-BMEC cells in Transwell assays. 6 In BBB models with primary BMECs (pBMEC) obtained from rat, bovine, or human brains, the predictive accuracy has been improved as a result [6][7][8] ; however, obtaining pBMECs requires sacrificing a large number of animals and large-scale culture is an issue due to the limited yield of BMECs from donors. Recent studies have developed protocols for BMEC induction using stem cells such as embryonic stem (ES) cells, induced pluripotent stem cells (iPSCs), and hematopoetic stem cells (HPSCs) with high proliferation ability, which show properties similar, but not identical to pBMECs. Lippmann et al., reported protocols for differentiation of BMECs from ES cells and iPSCs, and iPSC-derived BMEC (iBMEC) showed high barrier function (>1000 O · cm 2 ) as well as expression of EC markers, Tjs markers, glucose transporter 1 (Glut-1), and efflux transporters, 9,10 and coculture of human iBMECs with pericytes, neurons, and astrocytes further increases the barrier function. 11 Human HPSC-derived BMECs exhibit good predictive performance against human K p,uu,CSF , although the BBB barrier function (*200 O · cm 2 ) 12 is not as high as that of the human BBB in vivo. 13 Studies have compared the ability of in vitro and in silico models versus in vivo rodent models to predict BBB permeability to specific drugs 4,[14][15][16] ; however, few studies have succeeded in accurately predicting in vivo drug permeability in the human CNS. 12 Therefore, in this study, we therefore investigated whether an in vitro BBB model using iBMECs can predict human in vivo drug permeability in the CNS.
TEER measurement TEER was measured on D10 and D11 using an EVOM2 voltohmeter with ENDOHM-6 chambers (World Precision Instruments, FL). The TEER was recorded at the peak value. Membrane resistance was calculated as TEER (O · cm 2 ) = measured resistance value (O) · surface area (cm 2 ).
ICC staining ICC was performed using an Image-iT Kit (Invitrogen, MA) on D10. Briefly, iBMECs on permeable Transwells were washed with Dulbecco's phosphate-buffered saline iPSCs were seeded on D0 and cultured in unconditional medium from D1-D5 and in ECM (+RA, +bFGF) from D5-D8 to differentiate BMECs. On D8, BMECs were subcultured on Transwells in ECM (+RA, +bFGF) on the apical side and the medium was changed to ECM (+RA) on D9. For monoculture, either ECM (+RA) or neuron medium was used in basolateral wells. For coculture, astrocytes and neurons were cultured in astrocyte/neuron medium (1:1) from D4-D8 and pericytes were seeded on the back side of the Transwell in pericyte medium on D7. BMECs were seeded on the upper side of the Transwells with pericytes and cultured with astrocytes/neurons on the basolateral side with either ECM or neuron medium from D8-D11. TEER was measured on D10 and D11, ICC was performed on D10, and drug permeability tests and mRNA were performed on D11. (B) Structures of hiBBB models. ECM is used in apical wells in all monoculture and coculture models. For Mono-ECM and Co-ECM, ECM was used in the basolateral wells, and for Mono-N and Co-N, neuron medium was used in the basolateral wells. bFGF, basic fibroblast growth factor; BMEC, brain microvascular endothelial cell; ECM, endothelial cell medium; hiBBB, human iPSC-derived blood-brain barrier; ICC, immunocytochemistry; iPSC, induced pluripotent stem cell; RA, retinoic acid; TEER, transendothelial electrical resistance.

Caco-2 and ratBBB kits
To compare drug permeability with non-BMEC and non-human cells, we performed drug permeability tests using Caco-2 human epithelial colorectal adenocarcinoma cells and rat pBMECs. A Caco-2 kit (KAC Co. Ltd., Japan) was used according to the manufacturer's instructions. Briefly, the medium was refreshed and the kits were incubated overnight and used for permeability tests the next day.
RatBBB kits (RBE-12; PharmaCo-Cell Company Ltd., Japan) were maintained according to the manufacturer's instructions. Briefly, the kits were thawed by adding thawing solution and subsequently cultured with Culture medium 1 (500 lL in apical well and 1500 lL in basolateral well). After overnight incubation, the medium was replaced with Culture medium 2 and the cells were maintained for 4 days before use in drug permeability tests. Both drug permeability tests using Caco-2 and ratBBB kits were performed following the protocol described for drug permeability tests in the Material and Methods section.

Drug transporter array
All reagents and array plates were purchased from Qiagen-SABioscience. Cells were extracted with lysis buffer on D11, homogenized with a QIAshredder, and mRNA was purified using an RNeasyPlus Micro Kit with DNaseI. Transcription of cDNA from mRNA was performed using an RT 2 First Strand Kit, and human drug transporter array (RT 2 Profiler PCR Array Human Drug Transporters, 330231 PAHS-070Z) was performed with RT 2 SYBR Green qPCR Mastermixes. All procedures were performed according to the manufacturer's instructions. Real-time polymerase chain reaction (PCR) was performed on a ViiA 7 Real-Time PCR System, and data analysis was performed using R statistical software (version 3.5.2). Heat map data were drawn based on DCt values normalized to the average of eight housekeeping genes (i.e., Tap1, Tap2, Vdac1, Vdac2, B2m, GAPDH, Hprt1, and RPLPO) and Ward's clustering method was used for the analysis.
Drug permeability tests Cells were incubated with Hank's Balanced Salt Solution (HBSS) (Wako) for 2 h on D11 and the TEER was measured before carrying out permeability tests. hiBBB models with barrier function less than 500 O · cm 2 were excluded from drug permeability tests. For Caco-2 and ratBBB kits, all Transwells were used for permeability tests. Drugs at a concentration of 10 lM (Tables 2 and 3) diluted in HBSS were added to apical wells for A-to-B samples or basolateral wells for B-to-A samples; the cells were then incubated at 37°C with agitation and 50-lL samples from the basolateral wells (A-to-B samples) and apical wells (B-to-A samples) were collected at 60 and 80 min. Drug concentrations were measured by LC/MS analysis (LC: Prominence, MS: LCMS-2010EV; Shimazu Co., Japan), and LC/MSMS analysis (LC: Acquity; Waters Co., MA; MS: TSQ QUANTIVA; Thermo Fisher Scientific, Inc.) was used for digoxin measurement. For the  data analyses, Papp (apparent permeability coefficients; cm/sec), and efflux ratios were calculated using the following formulas: where dQ/dT is the amount of drug transported per unit time; A is the membrane surface area; and C0 is the donor concentration at time 0.

Statistics
Comparisons of more than three groups were performed using two-way analysis of variance (ANOVA) followed by Tukey's-Kramer test. For TEER measurement, two-way repeated measures ANOVA followed by Bonferroni-Dunn test was used. Statistical significance was expressed as *p < 0.05 and **p < 0.01. Statcel3 software (OMS Publishing, Inc., Japan) was used for all statistical analyses. All values are presented as meanstandard deviation, unless otherwise indicated.

Drug transporter expression
Transporter expression profiles of iBMECs (n = 3) were compared with pBMECs (n = 1), Caco-2 cells (n = 1), iPSCs (n = 2), and human umbilical cord blood cells (HUVECs, n = 1). In contrast to HUVECs, iPSCs, and Caco-2 cells, the hiBBB models showed very similar expression profiles to that of pBMECs. In comparison to hPPP (Mono-hECM) and bPPP (Mono-bECM), pBMEC showed a closer expression profile to Mono-hECM than to Mono-bECM. There was almost no difference in transporter expression among the hiBBB models. With respect to typical efflux transporters, Caco-2 cells clearly showed higher expression of P-gp/ ABCB1 versus other cell types, while ABCC1/BCRP expression was almost the same among pBMECs, hiBBB models, and Caco-2 cells, and ABCG2/MRP-1 expression was the same between pBMEC and hiBBB models, but lower in Caco-2 cells (Fig. 3).  Table 2, which have been reported to be P-glycoprotein (P-gp) substrates. [19][20][21] Caffeine was used as a positive control for simple diffusion across the BBB and it showed high permeability in all in vitro BBB models. Mono-ECM showed a remarkably high efflux ratio in vinblastine and digoxin, and showed <2.0 efflux ratio in quinidine, colchicine, and glibenclamide. Co-N showed >2.0 high efflux ratio in glibenclamide, vinblastine, and digoxin, and showed <2.0 efflux ratio in quinidine and colchicine. Caco-2 cells showed high efflux activity for all P-gp substrates (Fig. 4A). We subsequently investigated the correlation between in vivo human drug permeability (Table 3, permeability test 2) and drug permeability of iBMECs (Mono-ECM and Co-N) compared with ratBBB and Caco-2 kits. hiBBB models showed a greater correlation with in vivo data compared with Caco-2 kit, with an efflux ratio greater than two for gabapentin, verapamil, and thioridazine, which are reported to have low CNS permeability (Mono-ECM, R 2 = 0.49; Co-N, R 2 = 0.60; Caco-2 kit, R 2 = 0.41). RatBBB kit showed higher correlation than hiBBB (ratBBB kit, R 2 = 0.73); however, high concentration of verapamil and gabapentin penetrated to basolateral well across ratBMECs, which indicates that the ratBBB kit is inappropriate as a drug permeability prediction tool in substrates for efflux transporter. Taken together, these results suggest that P-gp transporter expression is insufficient in hiBBB models; however, hiBBB models have a better ability to predict drug permeability compared with other prediction models.

Discussion
Although some reports have indicated benefits of using iBMECs for in vitro BBB modeling, 12,17 the degree of similarity between iBMECs and pBMECs is not well known. We therefore focused on the ability of hiBBB models to predict in vivo drug permeability, and the transporter expression profiles of iBMECs compared with pBMECs. We first confirmed that iBMECs express BMEC markers (i.e., Tjs, EC marker, and influx/efflux transporters). In addition, this study demonstrated that neuron medium improves barrier function without coculturing, and that hPPP improves the reproducibility of iBMEC differentiation. Given that PPP contains angiogenic factors such as vascular endothelial growth factor, bFGF, platelet-derived growth factor, angiopoietin, and epidermal growth factor, the robustness of BMEC differentiation might be affected by PPP's angiogenic potential, 22 as suggested by the difference between hPPP and bPPP. Indeed, in drug transporter array, pBMEC had a closer expression profile to Mono-hECM than to Mono-bECM, which indicates that the formulation of PPP is crucial for controlling BMEC function. Using the method of Lippoman et al., we succeeded in shortening the differentiation period and improving the robustness of BMEC differentiation.
Transporter expression profiles in iBMECs were similar to pBMECs, while the expression of some transporters was remarkably higher or lower in iBMECs. Higher expression in iBMEC was mostly observed for the SLC2 family of glucose transporters (i.e., SLC2A1 and SLC2A3) and the SLC22 family, which transports organic anions/cations (i.e., SLC22A3 and SLC22A8) 23 ; however, there was no consistent pattern. P-gp on BMECs is a gatekeeper protein for xenobiotics 24 and serves as an essential transporter to protect the CNS; however, P-gp expression was low in iBMECs, even though ICC clearly demonstrated P-gp expression. In iBMECs, the insufficient expression of P-gp is one of the features that needs to be improved upon to accurately predict the permeability of P-gp substrates. Indeed, hiBBB models (e.g., Mono-ECM and Co-N) fail to exert efflux activity with some P-gp substrates such as quinidine, colchicine, and glibenclamid. 19,21,25 In contrast, these models showed sufficiently high efflux activity in digoxin, verapamil, and vinblastine, which are known to be P-gp substrates. Since the transporter expression profiles of in vivo BMECs remain unclear, 26 it is important to more thoroughly understand the transporter expression of in vivo BMECs, and manipulating the transporter expression of iBMECs to be comparable to that of in vivo BMECs is crucial for developing in vivo-like hiBBB models.
To assess the ability of in vitro BBB models to predict drug permeability, we selected drugs for which in vivo pharmacodynamics in the human CNS have been previously established. 27 Both ratBBB and Caco-2 kits showed poor predictive ability, while hiBBB models For topiramate, bupropion, lamotrigine, and tacrine, all BBB models showed similar drug permeability. Mono-ECM and Co-N models showed high efflux activities for gabapentin, verapamil, and thioridazine and had a stronger correlation with human in vivo permeability. RatBBB and Caco-2 kits had low efflux ratios with gabapentin and verapamil, which have low central nervous system permeability, and correlated poorly with human in vivo data.
(both monoculture and coculture) correlated with human in vivo permeability. Drugs with high BBB permeability, such as topiramate, bupropion, lamotrigine, and tacrine, showed the same permeation ratios among the groups tested. Although transporters for topiramate, bupropion, lamotrigine, and tacrine have been identified as P-gp, dopamine transporter, 28 organic cation transporter 1 (OCT1), 29 and choline transporter, 30 respectively, not all transporters for these drugs have been determined. In this study, their high distribution in the CSF, as shown in in vivo studies, 27 was confirmed by the high permeability observed in the hiBBB models. Gabapentin is a substrate for SLC7A5 influx transporter, 31 which is known to localize at the BBB; however, based on in vivo studies, it shows poor distribution in the CNS. 32 The drug transporter array showed, in iBMECs, the extremely high SLC7A5 expression, while the permeability of gabapentin was poor in the hiBBB model employing iBMECs. Thus, transporter expression and in vivo permeability do not necessarily correspond. Verapamil is widely known as an efflux compound by Pgp, 33 and hiBBB models exhibited efflux activity for verapamil, while both ratBBB kit and Caco-2 kits showed low efflux activity for verapamil. Taken together, these results clearly indicate that drug permeability in hiBBB models connote mechanisms implicating efflux transport of gabapentin, verapamil, and thioridazine.

Conclusion
The hiBBB model is more reliable for predicting drug permeability compared with non-human and non-BMEC BBB models. Further elucidation of transport mechanisms by the BBB is essential for improving predictive accuracy.