In Silico Approaches for Studying Transporter and Receptor Structure

Home , Deltorphin, DPDPE

IN SILICO APPROACHES FOR STUDYING TRANSPORTER AND RECEPTOR

STRUCTURE-ACTIVITY RELATIONSHIPS

DISSERTATION

Presented in Partial Fulfillment of the Requirements for the

Degree Doctor of Philosophy in the Graduate School of

The Ohio State University

Cheng Chang, B.S.

****

The Ohio State University

2005

Dissertation Committee: Approved by Dr. James Dalton, Advisor

Dr. Peter Swaan ______

Dr. Mark Foster Advisor

Dr. Charles Daniels Graduate Program in Biophysics ABSTRACT

Transporter proteins and receptors play a pivotal role in drug absorption, distribution and excretion. However, very few of the transporters have been crystallized and not all pharmaceutically significant receptors have been studied extensively.

Nonetheless, currently available functional as well as structural data provide an attractive scaffold for generating combined models that merge ligand-based structure-activity relationship and protein-based homology structures. The resultant models offer features that extend the predictive function of previous single models. This dissertation is aimed at presenting alternative approaches for studying transporter and receptor structure by applying in silico technologies with the following specific aims: (1) to develop thoroughly validated, highly predictive Quantitative Structure Activity Relationship

(QSAR) models and pharmacophore models for pharmaceutically important transporters and receptors; (2) to generate comparative three-dimensional models for essential drug targets; and (3) to identify novel inhibitors towards significant drug targets through database screening using pharmacophore models generated in aim 1.

Chapter 1 presents an overview of in silico approaches for studying drug targets.

A summary of the significance of transporters and receptors in human health is provided

ii along with a comprehensive review of recent successful in silico applications. Also included is a detailed description of the methods used in later studies.

Chapter 2 – 9 describe the QSAR and pharmacophore studies as well as pharmacophore-based database screening results for transporters involved in: drug absorption, i.e., nucleoside transporter and peptide transporters; drug elimination, i.e., organic cation transporter, organic anion transporting polypeptides and drug efflux, i.e.,

P-glycoprotein, and for pharmaceutically important receptors, i.e., androgen receptor, bile acid receptor. The significance of each drug target is first presented, followed by the description of the modeling study. The implication of each model is discussed after the validation process. The database screening results are also listed with experimental verification, when available.

Chapter 10 will summarize previous studies and compare advantages and disadvantages of different in silico methods. It will also discuss future directions of in silico modeling studies based on the work outlined in this dissertation.

iii

DEDICATION

To my wife

ACKNOWLEDGMENTS

I would like to extend my gratitude to my advisor, Dr. Peter Swaan, for scientific guidance, encouragement and financial support throughout my graduate study, and for his patience in polishing my writing.

I am grateful to Dr. Sean Ekins for inspiring discussions and for reviewing various aspects of this dissertation.

I would like to thank my committee members, Dr. James Dalton, Dr. Mark Foster,

Dr. Charles Daniels, for their time and valuable suggestions. I would also like to thank

John Ohrn from Accelrys for making Catalyst software available for my graduate study.

I am indebted to all my lab mates, especially Mitch Phelps, without whose countless trips to the graduate school, my graduation would not have been possible.

Many thanks to Susan Hauser, Kathy Brooks, and Trudy Robinson. It is their timely help that made my graduation possible.

I am grateful to my parents for their unconditional support; to my wife, Xiaoyan, without whom my graduate life would have been much more boring.

VITA

Sept 9, 1975……………………… Born – Huhehot, P. R. China

July 1999 ………………………….. B.S. Biochemistry and Molecular Biology

University of Science and Technology of China

Sept 1999 – May 2003 …………… Graduate Research Assistant

The Ohio State University

May 2003 – present……………...… Graduate Research Assistant

University of Maryland at Baltimore

PUBLICATIONS

1 Banerjee, A., Ray, A., Chang, C. and Swaan, P.W. (2005) Probing the Role of Cysteine Residues in the Apical Sodium - Dependent Bile Acid Transporter (SLC10A2) Using Novel MTS - Bile Acid Conjugates. Biochemistry In Press May 26, 2005

2 Chang, C., Pang, K.S., Swaan, P.W. and Ekins, S. (2005) Comparative Pharmacophore Modeling of Organic Anion Transporting Polypeptides: A Meta- analysis of Rat Oatp1a1 and Human OATP1B1. J Pharmacol Exp Ther In Press April 21, 2005

3 Chang, C., Ray, A. and Swaan, P.W. (2005) In Silico Strategies for Modeling Membrane Transporter Function. Drug Discovery Today 10 (9), 663-671

4 Ekins, S., Johnston, J.S., Bahadduri, P., D'Souza, V.M., Ray, A., Chang, C. and Swaan, P.W. (2005) In Vitro and Pharmacophore-Based Discovery of Novel hPEPT1 Inhibitors. Pharm Res 22 (4), 512-517

5 Suhre, W.M., Ekins, S., Chang, C., Swaan, P.W. and Wright, S.H. (2005) Molecular determinants of substrate/inhibitor binding to the human and rabbit renal organic cation transporters hOCT2 and rbOCT2. Mol Pharmacol 67 (4), 1067-1077

6 Bohl, C.E., Chang, C., Mohler, M.L., Chen, J., Miller, D.D., Swaan, P.W. and Dalton, J.T. (2004) A ligand-based approach to identify quantitative structure- activity relationships for the androgen receptor. J Med Chem 47 (15), 3765-3776

7 Chang, C., Swaan, P.W., Ngo, L.Y., Lum, P.Y., Patil, S.D. and Unadkat, J.D. (2004) Molecular requirements of the human nucleoside transporters hCNT1, hCNT2, and hENT1. Mol Pharmacol 65 (3), 558-570

8 Yates, C.R., Chang, C., Kearbey, J.D., Yasuda, K., Schuetz, E.G., Miller, D.D., Dalton, J.T. and Swaan, P.W. (2003) Structural determinants of P-glycoprotein- mediated transport of glucocorticoids. Pharm Res 20 (11), 1794-1803

9 Zhang, E.Y., Phelps, M.A., Banerjee, A., Khantwal, C.M., Chang, C., Helsper, F. and Swaan, P.W. (2004) Topology scanning and putative three-dimensional structure of the extracellular binding domains of the apical sodium-dependent bile acid transporter (SLC10A2). Biochemistry 43 (36), 11380-11392

10 Zhang, E.Y., Phelps, M.A., Cheng, C., Ekins, S. and Swaan, P.W. (2002) Modeling of active transport systems. Adv Drug Deliv Rev 54 (3), 329-354

FIELDS OF STUDY

Major Field: Biophysics

Computer Aided Drug Design

vii

TABLE OF CONTENTS

ABSTRACT...... ii

DEDICATION...... iv

ACKNOWLEDGMENTS ...... v

VITA...... vi

LIST OF FIGURES ...... xii

LIST OF TABLES...... xvi

CHAPTER 1 INTRODUCTION ...... 1

1.1 Receptor-based methods ...... 3

1.1.1 Comparative modeling...... 4

1.1.2 Recent applications ...... 5

1.2 Ligand based methods...... 10

1.2.1 Pharmacophore modeling ...... 11

1.2.2 3D-QSAR modeling...... 16

1.2.3 Recent applications ...... 22

1.3 Significance of this dissertation work...... 26

CHAPTER 2 MOLECULAR REQUIREMENTS OF THE HUMAN NUCLEOSIDE

TRANSPORTERS HCNT1, HCNT2, AND HENT1...... 33

2.1 Introduction...... 33

2.2 Materials and Methods...... 35

viii

2.3 Results and Discussion ...... 38

2.4 Conclusion ...... 49

CHAPTER 3 IN VITRO AND PHARMACOPHORE BASED DISCOVERY OF

NOVEL HPEPT1 INHIBITORS ...... 63

3.1 Introduction...... 63

3.2 Methods...... 65

3.3 Results and Discussion ...... 68

CHAPTER 4 MOLECULAR DETERMINANTS OF SUBSTRATE/INHIBITOR

BINDING TO THE HUMAN AND RABBIT RENAL ORGANIC CATION

TRANSPORTERS HOCT2 AND RBOCT2...... 78

4.1 Introduction...... 78

4.2 Materials and Methods...... 80

4.3 Results...... 85

4.4 Discussion...... 94

CHAPTER 5 COMPARATIVE PHARMACOPHORE MODELING OF ORGANIC

ANION TRANSPORTING POLYPEPTIDES: A META-ANALYSIS OF RAT

OATP1A1 AND HUMAN OATP1B1 ...... 115

5.1 Introduction...... 115

5.2 Methods...... 118

5.3 Results...... 120

5.4 Discussion...... 125

CHAPTER 6 STRUCTURAL DETERMINANTS OF P-GLYCOPROTEIN-

MEDIATED TRANSPORT OF GLUCOCORTICOIDS ...... 144

6.1 Introduction...... 144

6.2 Methods...... 147

6.3 Results...... 153

6.4 Discussion...... 158

CHAPTER 7 USING PHARMACOPHORES TO RAPIDLY IDENTIFY

MOLECULES WITH AFFINITY FOR P-GLYCOPROTEIN...... 172

7.1 Introduction...... 172

7.2 Materials and Methods...... 175

7.3 Results...... 177

7.4 Discussion...... 180

CHAPTER 8 A LIGAND-BASED APPROACH TO IDENTIFY QUANTITATIVE

STRUCTURE-ACTIVITY RELATIONSHIPS FOR THE ANDROGEN RECEPTOR194

8.1 Introduction...... 194

8.2 Materials and Methods...... 197

8.3 Results...... 202

8.4 Discussion...... 205

8.5 Conclusion ...... 209

CHAPTER 9 COMPARATIVE MODELING OF THE BILE ACID RECEPTOR .... 225

9.1 Introduction...... 225

9.2 Methods...... 226

9.3 Results and discussion ...... 228

CHAPTER 10 CONCLUSION...... 238

10.1 Summary and significance...... 238

10.2 Evaluation of in silico approaches ...... 242

10.3 Future perspectives ...... 243

BIBLIOGRAPHY...... 245

INDEX ...... 270

LIST OF FIGURES

Figure 1.1. General protocol for modeling of transporter and receptor proteins...... 28

Figure 2.1. Chemical structures of uridine analogs and drugs used in QSAR analyses... 55

Figure 2.2. Chemical structures of adenosine analogs and drugs used in QSAR analyses

...... 56

Figure 2.3. Chemical structures of cytidine analogs and drugs used in QSAR analyses . 57

Figure 2.4. Inhibition pharmacophore models for the three nucleoside transporters ...... 58

Figure 2.5. 3D-QSAR model of nucleoside transporter hCNT1 ...... 59

Figure 2.6. 3D-QSAR model of the nucleoside transporter hCNT2...... 60

Figure 2.7. 3D-QSAR model of the nucleoside transporter hENT1...... 61

Figure 2.8. Predictions for p38 MAPK inhibitors...... 62

Figure 3.1. HIPHOP Pharmacophore for hPepT1 substrates...... 76

Figure 3.2. Visualization of high scoring molecules discovered with and fitted to the hPepT1 ...... 77

Figure 4.1. Time course of TEA uptake into CHO cells stably transfected with either hOCT2 or rbOCT2...... 101

Figure 4.2. Kinetics of TEA transport in CHO cells stably transfected either hOCT2 or rbOCT2 ...... 102

xii

Figure 4.3. Effect of increasing concentration of several test inhibitors on uptake of

[14C]TEA mediated by either hOCT2 or rbOCT2...... 103

Figure 4.4. Relative effect on increasing concentrations of ephedrine...... 104

14 Figure 4.5. Comparison of IC50 values for inhibition of [ C]TEA transport mediated by hOCT2...... 105

Figure 4.6. Comparison of the relative effect of each test compound as an inhibitor of

hOCT2 versus rbOCT2...... 106

Figure 4.7. Relationship between the IC50 for inhibition of TEA transport mediated by hOCT2 and that compound’s calculated oil:water partition coefficient...... 107

2 Figure 4.8. Correlation between Cerius predictions and actual LogIC50...... 108

Figure 4.9. Structural overlap of the OCT2 compounds and CoMFA contour maps.... 109

Figure 4.10. Relationship between CoMFA predicted and measured LOGIC50 ...... 110

Figure 4.11. HIPHOP pharmacophores for rbOCT2 and hOCT2 ...... 111

Figure 5.1. OATP Pharmacophores...... 138

Figure 5.2. Test set predictions for the rat Oatp1a1-CHO pharmacophore...... 140

Figure 5.3. Mapping of DHEAS to three pharmacophore models ...... 141

Figure 5.4. Structural formulas of the compounds used in this study...... 143

Figure 6.1. Chemical structures of glucocortocoid receptor substrates...... 165

Figure 6.2. Calcein formation (dCAL/dt) in L-MDR1 and LLC-PK cells ...... 166

Figure 6.3. Transepithelial transport of cortisol...... 167

xiii

Figure 6.4. Teff and Peff versus PC...... 168

Figure 6.5. Pharmacophore model derived from P-glycoprotein ...... 169

Figure 6.6. CoMFA and COMSIA models for glucocorticoid analogs...... 170

Figure 6.7. Experimental versus predicted logTeff plot of the final CoMFA...... 171

Figure 7.1. Database scoring metrics...... 190

Figure 7.2 Alignment of test molecules to the P-gp pharmacophore model ...... 192

Figure 7.3. Venn diagrams showing the overlap of known P-gp...... 193

Figure 8.1. Alignment points for each set of compounds...... 219

Figure 8.2: Plot of residuals and predictions ...... 220

Figure 8.3. CoMFA model and ligand interactions ...... 221

Figure 8.4. CoMFA contours at different levels for AR...... 222

Figure 8.5. Overlap of a S-4, DHT, and a tricyclic quinolinone from the docked conformation into the AR homology receptor...... 223

Figure 8.6. Overlap of AR crystal structure...... 224

Figure 9.1. Initial machine generated multiple sequence alignment among 1ie9A, 1fm9D and BAR...... 232

Figure 9.2. Final manually modified multiple sequence alignment among 1ie9A, 1fm9D and BAR...... 233

Figure 9.3. Ramachandran plot of BAR LBD comparative model and the comparison of model and templates ...... 234

xiv

Figure 9.4. MatchMaker energy plot ...... 235

Figure 9.5. FlexX docking results...... 236

Figure 9.6. Comparison of model result with crystal structure...... 237

LIST OF TABLES

Table 1.1. Summary for membrane transporter comparative modeling studies...... 30

Table 2.1. Intramolecular atomic distances between pharmacophoric feature points ...... 52

Table 2.2. Statistical parameters for 3D-QSAR analyses...... 53

Table 2.3. QSAR predictions of test compounds (CoMFA)...... 54

Table 3.1. Catalyst CMC database search results...... 73

Table 3.2. Hit list from hPepT1 ...... 74

Table 4.1. Observed IC50 values for hOCT2 and rbOCT2 for structurally diverse organic cations ...... 112

Table 4.2. IC50 values for the inhibition of TEA by the phenylpyridiniums and

quinoliniums ...... 113

Table 4.3. CoMFA statistics for human and rabbit OCT2 models ...... 114

Table 5.1. Literature human Km data for OATP1B1 ...... 130

Table 5.2. Literature Km data for rat Oatp1a1...... 131

Table 5.3. Summary of the molecules tested with human OATP1B1 and rat Oatp1a1 . 133

Table 5.4. Model building and scrambling (trial average) summary...... 134

Table 5.5. OATP pharmacophore coordinates...... 135

Table 6.1. Kinetic parameters of glucocorticoid transepithelial transport...... 163

Table 6.2. Relative intramolecular distances between pharmacophoric feature points.. 164

xvi

Table 7.1. Coordinates for inhibitor pharmacophore 2...... 183

Table 7.2. Model statistics for the inhibitor model 1...... 184

Table 7.3. The metrics of the P-gp...... 185

Table 7.4. The metrics of the merged Maybridge database screening hit list ...... 186

Table 7.5. The predictions for NSC compounds...... 187

Table 7.6. Hit list from SCUT database search with inhibitor model 2...... 188

Table 7.7. Average properties of each section...... 189

Table 8.1. Structures and predictions of AR CoMFA training set...... 210

Table 8.2. Test set structures and predictions...... 217

Table 8.3. Statistics for CoMFA model...... 218

Table 9.1. Flexible docking results of different cholic acids...... 231

xvii

CHAPTER 1

INTRODUCTION

As a general introduction to this dissertation, the first chapter will give an overview of the current challenges associated with transporter studies that may be addressed by in silico methods, i.e., computational approaches that have generally been applied to receptor systems in the past. A detailed review of different in silico methods,

as well as their recent successful applications, is presented to emphasize the significance

of this dissertation work.

Transporters are polytopic membrane proteins that are indispensable to the

cellular uptake and homeostasis of many essential nutrients. Furthermore, it is now

known that many drugs are transporter substrates, thus substantiating their intricate

involvement in all facets of drug absorption, tissue distribution, excretion and toxicity as

well as drug pharmacokinetics and pharmacodynamics – all widely accepted critical

parameters for a new drug to survive the drug discovery and development pipeline.

Clearly, a detailed understanding of transporter structure, function and mechanism would

1 have large potential value in drug discovery. Despite the significance of membrane transporters in drug discovery, most are poorly characterized at the atomic level due to difficulties associated with expressing (1) and crystallizing these proteins (2). To date there are only 86 membrane protein structures available (see http://blanco.biomol.uci.edu/Membrane_Proteins_xtal.html for an up-to-date listing), of which only two are from the major facilitator and three from the ATP-binding cassette

(ABC) superfamilies. As a result, our knowledge of transporter structure and mechanism has lagged far behind the insight into their biochemical and functional properties. In silico techniques are a useful tool to circumvent the difficulties associated with traditional crystallization techniques and fill the gap between our knowledge of transporter structure and transporter protein properties. This trend is confirmed by an increasing number of publications that simulate transporter structure and mechanism using singular as well as combinatorial in silico approaches; these models are consecutively validated by empirical methods.

In silico approaches are not limited to transporter studies where the target structure is unknown. They have been widely applied to receptors, the quintessential drug target. When the receptor crystal structure is unavailable, as exemplified in the bile acid receptor study (chapter 9), in silico methods play a pivotal role in gaining an understanding of receptor structure-activity relationships. Even when a receptor three- dimensional structure is available as illustrated by the androgen receptor study (chapter 7), more convincing QSAR models can be derived and model interpretation is more straightforward.

In silico models are not limited to inferring substrate binding interactions. As a

generally cost-effective method, in silico screening has the advantage of being able to

significantly reduce experimental efforts in the screening phase of drug discovery (3).

This is exemplified by the hPepT1 study (chapter 3) and P-gp transporter study (chapter

6).

The following sections will present a comprehensive overview of applications in

ligand-based methods such as pharmacophore modeling and quantitative structure-

activity relationships (QSAR) in which transporter substrate specificity is analyzed in the

absence of protein structure as well as the receptor-based method in which the transporter

or receptor structure is directly modeled to gain insight into the transport process. The

different approaches are illustrated in Figure 1.1.

1.1 Receptor-based methods

The primary amino acid sequence of many transporters is known and their

secondary structures can be solved in a relatively straightforward fashion using

bioinformatics tools such as hydropathy plots and, preferably, by a combination of

experimental verification methods such as N-glycosylation analysis or epitope insertion scanning (4, 5). Although only a few transporter proteins have yielded to X-ray crystallographic analyses resulting in high-resolution three-dimensional information, new structures are emerging at a steady rate. This opens the opportunity for generating homology or comparative models. In general, when two proteins have adequate sequence identity (in general > 60%), and the experimentally determined 3D structure of one of

3 these proteins is known, comparative protein models may be constructed, even when the two proteins are not functionally related (6).

1.1.1 Comparative modeling

The procedure of comparative modeling is best illustrated with the program

Modeller (7), which was specifically developed by Sali and colleagues for this purpose.

First, possible structural templates as well as related sequences are retrieved using the query sequence. All template structures are first aligned to generate a sequence block, which is subsequently aligned to all related sequences of unknown 3D structure.

Extensive manual adjustment of the computer generated multiple sequence alignment is required to ensure alignment of conserved residues as well as to move gaps to the loop regions that connect the transmembrane domains (TMD). This is particularly important because most amino acid variability is observed in the extramembranous loop regions.

Importantly, high homology should not be the sole criterion for manually adjusting the alignment as this might not correspond to high amino acid identity. One or more templates are selected based on sequence similarity and phylogenetic analysis. Modeller will generate several models based on the alignment which are in turn evaluated by internal self-consistency checks and external programs that check structural protein quality, such as ProCheck (8) or WHATIF (9). The cycle of template selection, sequence alignment adjustment, modeling and evaluation are repeated until no further improvements to the model are observed.

1.1.2 Recent applications

Comparative modeling techniques have allowed the elucidation of 3D structures for a large number of transporters that would otherwise not be available to researchers

(Table 1.1). This area has accelerated significantly after the recent publication of high- resolution crystal structures for two bacterial transporters, the LacY lactose permease

(3.5Å) (10) and the GlpT glycerol-3-phosphate transporter (3.3Å) (11). Both structures represent polytopic membrane proteins comprising 12 transmembrane segments.

Importantly, LacY and GlpT are functionally different transporters, but share a similar folding pattern. This has encouraged researchers to speculate that all Major Facilitator

Superfamily (MFS) transporters share folding patterns, thus initiating novel opportunities for comparative modeling and we can expect these models to appear in the literature soon.

For example, the group that crystallized GlpT recently described a comparative model of the human glucose-6-phosphate transporter (G6PT), which is involved in the pathology of glycogen storage disease type Ib (12). The authors identified two positively charged residues, Asp28 and Lys240, as part of the substrate-binding site. Disease causing 28 missense and 2 deletion mutations were mapped onto the comparative model to provide information on the transporter's molecular mechanism. Three different disease-causing mechanisms were proposed: mutations may modify the substrate-binding site, change the

N- and C-terminal domain interface, or destabilize the protein. These observations agree with the available experimental data on G6PT, providing further support for the comparative modeling approach (12).

Facilitative glucose transporter-1 (Glut1)

Due to its role in glucose transport and diabetes, Glut1 is perhaps the most

extensively studied membrane transporter, yet an atomic level structural description is

still absent. Several in silico models have been generated to aid in explaining the

available experimental data. For example, Zuniga and colleagues (13) constructed a 3D

Glut1 model using a "piecemeal strategy"; experimentally derived information was used

to guide the step-by-step model building, supplemented by comparative modeling of

extramembranous loops and molecular dynamics (MD) optimization of the overall

protein. MD simulations are ideally suited to allow the protein to relax in a simulated

membrane environment. The final model assisted in explaining experimental data not

used in model building, including the observation of interaction between Asn288 and glucose during a MD run, which explained the observed 10-fold reduction in glucose transport of the mutant Asn288Cys (14). With the availability of the crystal structure for

LacY, a potential template for Glut1, a comparative model was generated (15) to further

illustrate their mutagenesis results. The same group continued to verify and optimize their

model using new experimental results (16, 17). This exemplifies a new trend of feedback

between in silico studies and bench experiments, i.e., the in silico models help to

illustrate and explain experimental discoveries, while experimental results aid in

continuously optimizing theoretical models.

ATP-binding cassette (ABC) transporters

The ABC superfamily of transporters have been extensively studied due to their

role in drug efflux, drug resistance and drug-drug interactions. As a result, a wealth of

6 experimental information is available. To date, all three available full-length ABC transporter crystal structures are bacterial in origin. In order to further our understanding of human ABC transporter mechanisms, however, a human ABC transporter atomic structure is pivotal. The most widely studied ABC transporter, P-glycoprotein (P-gp), was initially modeled using the crystal structure of a second ABC protein, the E. coli lipid A transporter (MsbA) (18). This template shares approximately 30% homology with P-gp.

It should be noted that, due to the relatively low resolution (4.5Å), the original MsbA crystal structure file does not contain coordinates for TMD side chain or backbone atoms

(except C-alpha) and does not account for 39 out of 582 amino acid residues. In fact, the

C-alpha coordinates of the nucleotide binding domain (NBD) are only partially available

(60%). As a result, special emphasis was applied to modeling the missing backbone and side-chain atoms. Overall, the resultant model can only be used qualitatively, mainly due to the missing template information as well as alignment uncertainties. Regardless, the model provided additional information by revealing that the three intracellular domains of each transporter half have distinct and selective interactions with the core domain

(which contains the ATP binding site), the α-domain (i.e., helices α3 to α5) and the conserved Q-loop that links the previous two domains together. The authors also noticed highly conserved aromatic residues in the internal chamber of the protein, as well as a correlation of chamber cavity dimensions and substrate size and charge, which together suggest an important role of the internal chamber in drug transport.

Stenham and colleagues (19) challenged the validity of the E. coli MsbA crystal structure. By comparing the crystal structure with their disulfide cross-linking studies as well as other non-crystallographic experimental evidence, the authors suspected that the

orientation of E. coli MsbA TMD with respect to the NBD observed in the crystal

structure does not reflect the physiological association. Instead, they comparatively

modeled the P-gp TMD and NBD according to E. Coli MsbA independent from each

other and subsequently assembled the protein based on cross-linking data. Additionally,

they incorporated structural data from another bacterial ABC transporter, BtuC (vitamin

B12 transport system permease), in particular the correct orientation of the NBD domains.

The final model of P-gp obtained contains a consensus NBD:NBD interface and a

parallel TMD:TMD interface that is structurally consistent with chemical cross-linking

data as well as previous electron microscopy data. The physiologically correct orientation

of the NBD with respect to the TMDs was again confirmed in the crystal structure of

MsbA in V. cholera (20).

The two diverging comparative models of P-gp were later reconciled following a

theory by Lee and co-workers (21), who proposed that P-pg could exist in two dynamic

conformations, an open conformation (18) and a closed conformation (19). The authors

also suggested that by rotating the two monomers toward each other, the E. coli MsbA

crystal structure (open) could be transformed into the closed conformation.

The V. cholera MsbA structure was used to model another major ABC transporter,

MRP1 (Multidrug Resistance Associated Protein 1) (22). The model implied that Phe594 forms a hydrophobic pocket with four previously identified residues that were shown to be important for activity. The hypothesis that Phe594 would also play a role in MRP1 transport activity was verified by site directed mutagenesis. This exemplifies that in silico models may guide experimental design to identify important functional amino acid residues, providing insight into function and mechanisms of biologically important

8 processes. Using the same template structure, Ecker and colleagues (23) built a comparative model for the bacterial efflux transporter LmrA, which is a close ortholog to human P-gp. The authors identified the drug-binding site using photoaffinity labeling and built a comparative model to illustrate their results and provide an overall transporter structure to guide further experimental design.

In general, the relatively low quality of the E. coli MsbA structure diminished its usefulness as a template for comparative modeling. With the assumption that molecular modeling could extend the usability of low-resolution crystal structures, Campbell and co-workers optimized the 4.5 Å MsbA crystal structure to atomic resolution by combining comparative modeling techniques and molecular dynamics simulation (24).

The authors regenerated the missing backbone and side chain atoms using

MAXSPROUT (25) and Modeller. The coordinates for protein gaps that had readily available templates (e.g. NBD) were generated by comparative modeling. Additional missing C-alpha positions were predicted by QUANTA

(http://www.accelrys.com/quanta/) and PSIPRED (26) resulting in a model with acceptable stereochemical quality parameters such as dihedral and torsional angles. The stability of the MsbA model was further verified by a 2ns MD simulation after the protein was inserted into a membrane-mimetic octane slab with full solvation on both sides.

During the MD simulation, the authors observed significant structural rearrangement in the TMDs, which is corroborated by previous evidence that repacking of the TMDs occurs (27) during the P-gp transport cycle. Based on their simulation data the authors suggested that the crystal structure of MsbA dimer might not correspond to the MsbA

dimer in vivo, being the result of crystal packing effects rather than the in vivo dimeric

state.

Protein-mediated transport is a dynamic process involving several intermediary

conformations. Thus, a static structure may not be adequate to study the molecular

mechanisms of transport and MD simulations could overcome these limitations. For

example, to determine the catalytic mechanism of the NBD of ABC transporters Jones

and colleagues performed a 390 ps MD simulation of the bacterial histidine permease

(HisP) NBD domain (28). The authors discovered that the peptide bond between Phe99 and Gln100 serves as the hinge point in ABC transporters, which moves the conserved

glutamine in and out of the catalytic site. They also identified key interfaces involved in

TMD and NBD communication that, together with previous experimental data, could be

translated to a detailed catalytic cycle of the ABC transporter superfamily.

1.2 Ligand based methods

The aforementioned in silico approaches based on molecular level transporter

models may be used to predict substrate and inhibitor binding modes, thus representing a

useful tool in the discovery of novel transporter ligands. Although comparative models

can be generated successfully for membrane proteins with 7 or 12 TMDs, this approach is

not yet feasible for proteins without a suitable TMD template. In these cases, techniques

that do not require knowledge of transporter structure, such as pharmacophore modeling

and 3D-QSAR modeling, can be applied. They are employed to correlate biological

activity with molecular descriptors or features that represent the interaction between the

ligand and the protein. The obvious assumption of this approach is that the biologically

active compounds have specific components represented by chemical features or

electrostatic and steric fields around the molecules that lead to the activity.

This approach has been applied successfully to generate both pharmacophore and

3D-QSAR models for P-glycoprotein (29, 30), organic cation transporters (31, 32), bile

acid transporters, nucleoside transporter (33), hPepT1 (34, 35), OATPs (36), as well as enzymes and receptors (37, 38). Table 1.2 represents a comprehensive overview of some current applications.

Several commercial tools are available to generate 3D-QSAR models. Based on the number of publications, the three most widely applied programs are perhaps comparative molecular field analysis (CoMFA) (39), comparative molecular similarity index analysis (CoMSIA; Tripos Associates, Inc.) (40) and Catalyst (Accelrys, Inc.).

After a thorough literature search these programs were reportedly used in 515, 113, and

92 publications, respectively, of which 31, 10, and 8 explicitly dealt with transporter models. These approaches are discussed in detail followed by recent studies involving pharmaceutically relevant systems.

1.2.1 Pharmacophore modeling

A pharmacophore is the representation of the spatial arrangement of structural features that are required for a certain biological activity. The model can give insight into the binding or inhibition process, form a basis for the design of more active compounds, and assist the identification of other molecules that contain the same pharmacophore.

Three automated programs are widely used for pharmacophore generation, DISCO

(DIStance COmparisons) (41), GASP (42) and Catalyst/HIPHOP (43). All programs

11 attempt to determine the common features based on the superposition of active compounds using different algorithms. They merely require the input of active molecules and no measured activity is needed. One program, Catalyst/HypoGen, uses a combination of QSAR and pharmacophore methods together (44). Instead of only the active compounds, Catalyst/HypoGen analysis requires a full range of test compounds from active to inactive along with their measured activities derived from experimental data.

The result extends the usual pharmacophore. It not only identifies a query compound as active or inactive like in the tradition of a pharmacophore model, but it also predicts the activity based on the regression of the training dataset. This represents the capability of a

QSAR model and is referred to as the HypoGen pharmacophore model in this dissertation.

All the above mentioned programs were applied during this thesis work. More recently, a new pharmacophore detection algorithm called DANTE has been developed (45).

However, this method is not discussed because there are few if any publications available.

DISCO

The DISCO method is based on the assumption that the pharmacological potency of a compound can be represented by its structural points of pharmacological interest, which are defined as "DISCO features", i.e., hydrophobic centers, hydrogen bond (H- bond) donors, H-bond acceptors, positive charges, and negative charges. The program utilizes a clique detection algorithm to find the maximum common graph that contains geometrical placement of structural points among all active compounds, which represents the pharmacophore model. To find the bioactive conformation instead of the minimum energy conformation for each molecule, a maximum of 25 conformers within an arbitrary

70.0 kcal/mol energy cutoff can be generated for each molecule using the MultiSearch function. Typically, DISCO features are assigned to all conformers. The reference compounds to which all other compounds are aligned should be the molecule with few features and few conformers (41). The feature distance tolerance is normally set at 1.0 Å and DISCO is initially run considering all the potential "feature" points. Additional runs with specification of the number of features based on chemical intuition can be also carried out subsequently. Other than providing insight into the transporter substrate binding process, the resulting pharmacophore models can be used as a useful routine to superimpose each set of the molecules and provide a starting point for 3D-QSAR analysis.

GASP

GASP (Genetic Algorithm Similarity Program) is a genetic algorithm, developed for the superimposition of sets of flexible molecules (42). It also assists the perception of pharmacophore models. When aligning a set of molecules, GASP attempts to optimize the orientation and conformation of molecules by quickly and efficiently fitting them to similarity constraints. Molecules in conformations and orientations that best fit the constraints are allowed to produce more offspring similar to themselves in the genetic algorithm. Each alignment of N molecules is encoded in a chromosome containing N binary strings and N-1 integer strings. A binary string encodes the conformations for every molecule. Each byte of a string encodes an angle of rotation about a rotatable bond.

The integer string encodes the mapping between features in a molecule to features, of the same type, in the reference molecule. All features in all molecules are used to compute the fitness function. Each integer string has a length L (number of features in the

13 reference molecule). One major limitation of GASP is that each run can only align four to five molecules. So a template should first be selected and all structures divided into groups of four or five compounds. Each group will be aligned to the same template.

When generating the alignment, to broaden the diversity of conformations that are considered, the “population size” is increased to 125 and the “allele mutate weight” to 96.

With increased population size and mutation rate, conformations with more diversity are generated. To avoid possible problems with convergence because of the increased mutation rate, the convergence criteria should also be loosened by increasing the “fitness increment” to 0.02. Normally, four to ten alignments are generated for each group and visual inspection will determine the best fit. Alignments from different groups are merged to generate the final overlap for subsequent QSAR studies.

Catalyst HIPHOP

A pharmacophore model is referred to as “hypothesis” in the Catalyst program to reflect the fact that generated models might not necessarily represent the true pharmacophore. The terms pharmacophore and hypothesis will be used interchangeably in the remainder of this dissertation. The HIPHOP module generates a pharmacophore based on common chemical features among active molecules (43). This method is different from DISCO in that all active molecules are treated as the reference compound once, based on which feature configurations are identified by a pruned exhaustive search which is calculated and expanded to other molecules, if possible. The results are ranked by how well the molecule fits to the proposed hypothesis as well as how simple the final hypothesis is. When generating hypotheses, customizations to feature requirement in

14 model generation are possible according to each situation so that not all molecules have to match all features in the proposed hypothesis.

The training set conformations are produced using up to 255 conformers with the best conformer generation method, allowing a maximum energy difference of 20 kcal/mol. This setting also applies to the HypoGen module described below. Up to five chemical features including customized features can be specified in each model generation process. Normally the chemical space can be covered by specifying hydrophobic, H-bond acceptor, H-bond donor, positive ionizable, and negative ionizable features.

Catalyst HypoGen

With the input of a full range of training set compounds ranging from inactive to active, the HypoGen algorithm is set to generate hypotheses with features common among active molecules and missing from the inactive molecules (44). This is accomplished in three steps, a constructive step, a subtractive step and an optimization step. Hypotheses common amongst the active compounds are identified in the constructive step, which is very similar to the regular pharmacophore perception procedure. Hypotheses common among inactive compounds are removed from the previous result in the subtractive step. The resultant hypotheses are then optimized using simulated annealing to further fine tune the model parameters thus improving the model quality. The ten simplest hypotheses that best correlate the estimated activities with measured activities are ranked and returned for further selection. A wide range of activity

(4 – 5 orders of magnitude) among the training set is preferred for a high correlation and

15 predictive pharmacophore. The quality of the structure activity correlation between the estimated and observed activity values can be estimated by means of an r value. The statistical significance of the retrieved hypothesis is verified by permuting (randomizing) the response variable ten times, i.e., the activities and structures of the training set compounds can be mixed ten times so that each value was no longer assigned to the original molecule. The Catalyst hypothesis generation procedure is then repeated for each of these ten training sets. The total energy cost of the generated pharmacophores can be calculated from the deviation between the estimated activity and the observed activity, combined with the complexity of the hypothesis (i.e., the number of pharmacophore features). A null hypothesis can also be calculated which presumes that there is no relationship in the data and the experimental activities are normally distributed about their mean. Hence, the greater the difference between the energy cost of the generated hypothesis and the energy cost of the null hypothesis, the less likely it is that the hypothesis reflects a chance correlation. A recent publication gives an exhaustive recommendation for the size of the cost difference (46).

1.2.2 3D-QSAR modeling

A 3D-QSAR model consists of a mathematical equation describing potency as a function of 3D interaction fields around aligned training set compounds. The relationship between the 3D spatial change of interaction field values and the experimentally observed variations in the target property is produced through multivariate statistical analyses. This technique is often applied to assist the design of more potent inhibitors or substrates with higher affinity. CoMFA is the method of choice in QSAR studies as it correlates the two

16 major interactions, i.e., electrostatic and steric interactions, with bioactivity through a partial least square (PLS) analysis (39). CoMSIA is a similar technique in that it replaces the original interaction field calculation function with Gaussian type functions to avoid extreme values in interaction fields (40). GOLPE (47) is an alternative method that performs multivariate analysis based on interaction fields generated by GRID (48).

CoMFA

CoMFA has proven to be an especially useful QSAR technique with considerable success in drug design and drug transport (49-52). It explains the gradual changes in observed biological properties by evaluating the electrostatic (Coulombic interactions) and steric (van der Waals interactions) fields at regularly spaced grid points surrounding a set of mutually aligned ligands for a specific target protein. In a typical experiment, the available experimental data set is first randomly divided into a training set and a test set.

In some studies where the experimental data are very scarce (for example see the OATP study, chapter 5), all available data are used as the training set and external data are used to test the models. The statistical algorithm, PLS, is used to correlate the field descriptors with biological activities. The PLS algorithm used to derive 3D-QSAR models is a variation of principal component regression in which the original variables are replaced by a small set of linear combinations thereof. The latent variables generated are used for multivariate regression, maximizing the communality of predictor and response variable blocks. The advantage of PLS over other statistical methods is its ability to handle multivariate regression analysis in cases where the number of independent variables is greater than the number of samples. This is particularly useful in CoMFA (and CoMSIA)

17 analyses where the number of columns, representing the electrostatic and steric values measured by the probe atom at each lattice intersection point, is far greater than the number of rows, representing each molecule and its biological activity data. Overall, PLS reduces the risk of chance correlations (53). Both fields are calculated using an sp3 hybridized carbon probe atom (+1 charge at 1.52-Å van der Waals radius) on a 2.0-Å spaced lattice, which extends beyond the dimensions of each compound in the sest by 4.0

Å in all directions. These parameters are customizable and the default was used for the majority of studies in this thesis since they represent the most common interactions. A cutoff of 30 kcal/mol ensures that no extreme energy terms will distort the final model.

The indicator fields (54) and H-bond fields (55) generated by the "advanced CoMFA" module are also included in the analysis. The H-bond field descriptors are set to zero at sterically prohibited points. Sterically allowed points close to H-bond acceptors are assigned a nominal steric potential while the points close to H-bond donors were assigned a nominal electrostatic potential. As a result, the steric field actually represents the acceptor component while the electrostatic field essentially indicates the donor component. After the generation of field descriptors, a factor analysis is performed to help understand the clustering of inhibitors as well as identify potential outliers. When available, the experimental standard deviation is used as a weighting factor in PLS analyses, and calculation time can be decreased by using sample-distance PLS. The predictive value of the models is evaluated first using leave-one-out cross-validation.

This calculation will also determine the optimum number of components used in later conventional (non-cross-validated) PLS runs. The cross-validated coefficient, q2, is calculated as follows:

2 ∑(Ypredicted − Yobserved ) 2 Y q = 1− 2 ∑(Yobserved − Ymean ) Y

where Ypredicted, Yobserved, and Ymean are the predicted, observed, and mean values

2 of the target property (percentage inhibition), respectively. (Ypredicted - Yobserved) is the predictive error sum of squares. The standard error of the cross-validated predictions is known as “press”, and the root mean square of the conventional analysis is known as “s”.

The model with the optimum number of PLS components, corresponding to the lowest predictive error sum of squares value, is typically selected for deriving the final PLS regression models. In addition to q2, the conventional correlation coefficient r2 and its standard error are calculated. A plot of predicted versus experimental activity is then used to identify potential outliers. This process is repeated until no further improvements in q2 or no outliers could be identified. Results from alternative descriptor fields, such as logP, dipole moment and Molconn-Z parameters which represent a wide range of topological indices of molecular structure (56), can be compared and the model with the highest q2 is normally reported. Typically, a contour map of standard coefficients enclosing the top

20% lattice points where the QSAR strongly associates changes in CoMFA field values with changes in inhibition is created for visualization.

The "predict properties" command in the QSAR module is used to predict the percentage inhibition of test compounds. Test compounds are sketched, aligned and optimized using the same procedure as the training set. The correct alignment is essential

for accurate predictions. For some studies, the test compound is very similar to the

training set compound and they may only differ by a side chain. In this case, a

modification of the original aligned training set compound is performed, followed by

optimization and prediction because the new test compound maintained the original

alignment.

CoMSIA

The CoMSIA algorithm is an extension of CoMFA methodology (40). Both methods are based on the assumption that changes in binding affinities of ligands are related to changes in molecular properties, represented by interaction fields. They differ only in the implementation of the fields. Instead of calculating the steric fields using a

Lennard-Jones potential and electrostatic fields using a Coulombic potential, CoMSIA uses a Gaussian type function to avoid the extreme values generated by Lennard-Jones and Coulombic functions. The Gaussian function also results in smoother, less fragmented surfaces in the final model representation.

2 q −αriq AF ,k ( j) = ∑Wprobe,kWike i

where:

A is the similarity index at grid point q, summed over all atoms i of

the molecule j under investigation.

Wprobe, k is the probe atom with radius 1 Å, charge +1, hydrophobicity +1,

H-bond donating +1, H-bond accepting +1.

Wik is the actual value of the physicochemical property k of atom i.

riq is the mutual distance between the probe atom at grid point q and atom i of the test molecule.

α is the attenuation factor, with a default value of 0.3, and an optimal value normally between 0.2 and 0.4. Larger vales result in a steeper Gaussian function, and a strong attenuation of the distance-dependent effects of molecular similarity.

GOLPE

Generating Optimal Linear PLS Estimations (GOLPE; version 4.5, Multivariate

Infometric Analysis, Perugia, Italy) is an alternative QSAR method that was used to validate independently the CoMFA results (47). It performs multivariate regression analysis on the interaction fields around the molecules generated by GRID19 (Molecular

Discovery Ltd., Oxford, England). The type of the field depends on the probe, which should be selected based on the type of the interested interaction. In this thesis, a phenolic hydroxyl (OH) probe is selected because it can offer both H-bond donor and H-bond acceptor as well as the hydrophobic interaction. The interaction energies are calculated at

1.0-Å spaced grid points, which extend to be 4.0 Å beyond the dimensions of all the molecules in the training set. The generated interaction fields are selected and analyzed by GOLPE. A principal component analysis is also performed to check the distribution of objects and variables. A conventional PLS is performed followed by a leave-one-out cross-validated PLS. A PLS plot (T-U plot) of the first component is generated to

21 determine the outliers, which are excluded in the next run. The iteration is generally stopped when there are no obvious outliers in the PLS plot or no significant improvements in the result. Smart region design (57) generates Voronoi polyhedra for

1000 seed variables around 1 grid unit region of each seed variable and then collapses the polyhedra within 2 grid units of each seed variable. This maintains the three dimensional information and simplifies the following variable selection process. Based on smart region design, variables that best span the multidimensional weight space are selected by

D-Optimal preselection. Fractional factorial design selection is carried out to further select only the most informative variables. The final conventional PLS is then performed followed by the cross-validated PLS. Coefficient contour maps enclosing the top 20% lattice points at which the QSAR strongly associates changes in GOLPE field values with changes in bioactivity are generated for each model.

1.2.3 Recent applications

Monoamine transporters

Monoamine transporters are represented by the dopamine transporter (DAT), serotonin transporter (SERT) and norepinephrine transporter (NET). They are ion- coupled secondary transporters with twelve membrane spanning domains. Owing to their importance in substance abuse and depression, they have been studied extensively through QSAR analyses. Recently, Kulkarni and colleagues generated a highly correlative (q2, 0.695) and predictive (r2, 0.75) CoMFA models for DAT using inhibition data from a set of 71 mazindol analogues (58). The model identified the importance of the relative orientation of the two hydrophobic groups and the placement of a heteroatom

of the mazindol analogues that could serve as a guiding principle in the design of novel

analogues with higher DAT binding.

Using a different training set of 76 benztropine analogues, the same group

generated additional CoMFA models for DAT (59). In this study, the diphenyl ether

moiety was found to be vital for binding of this particular class of compounds to DAT.

The finding that the N-8 position tolerated various substitutions was used to synthesize a

series of N-8 substituted benztropine derivatives for testing. These compounds showed

potent inhibition of dopamine uptake with reduced lipophilicity, a step forward towards

increasing bioavailability. They also showed excellent selectivity of DAT over SERT and

NET. Thus, this study provides a practical example of CoMFA-guided drug discovery of

novel, potent DAT inhibitors with favorable physicochemical properties. An interesting

next step would be a combined model encompassing both benztropine and mazindol

analogues, thereby providing coverage for a much broader chemical space.

Despite inference from these ligand-based studies, the DAT structure details still

remain elusive. In order to generate an atomic description of the transporter, Ravna and

colleagues generated a DAT comparative model based on the electron density projection

map of another ion-coupled secondary transporter that also possessed twelve

transmembrane domains, Na+/H+-antiporter (NhaA), as well as site directed mutagenesis data on DAT (60). They also docked cocaine and its analogue, (–)-2β-carbomethoxy-3β-

(4-fluorophenyl) tropane (CFT), to this model. Based on their model, the authors

suggested that Asp79, Tyr252 and Tyr274 are the primary cocaine binding residues. They further proposed a dopamine transport mechanism, which involves transmembrane helices 1, 3, 4, 5, 7 and 11. This study is unique in that a synergistic combination of

23 comparative modeling and ligand-based design was applied. Additional insight was gained into the monoamine transport process which may ultimately lead to better drug discovery for the treatment of substance abuse and depression.

Combined In Silico Approaches

In silico methods represent one of numerous useful tools to study the transport mechanism and substrate affinity requirements of membrane transporters. They efficiently fill the gap between our knowledge of atomic level transporter structural mechanisms and in vitro transporter properties derived by biochemical experiments. This type of hybrid approach may eventually lead to the discovery of safer and more efficient drugs by targeting transporters or, as in the case of efflux pumps, avoiding them for increased cellular permeability. Combining the results of an array of computational methods to provide insight into the complexities of transporter structure-activity relationships is rapidly becoming an important approach to study transporter function.

Despite the utility of in silico approaches to studying membrane transporters, investigators are cautioned to avoid overextending model interpretation or extrapolation.

For example, the quality of a comparative model is limited by the validity of sequence alignment and the quality of template structure. Sequence homology between transporter families is usually too low to confidently build a robust model. To overcome these limitations, investigators have opted to forego classical sequence homology rules that have been applied to soluble proteins and demanded >60% sequence identity for the construction of a reliable model and >80% identity for a high-quality model. Instead, combination approaches are now actively being developed that focus on shared

24 membrane topology, satisfactory alignment of transmembrane regions, and comparative modeling of extracellular protein segments. For such models to be acceptable, they have to be critically evaluated with biochemical data. On the other hand, QSAR models are devoid of information on substrate-receptor interactions at the molecular level, even though they could be used to infer such interactions. A recent development is the combination of 3D-QSAR models with comparative models to extract simultaneous information from substrate affinity data as well as atomic level binding site models (61).

Each technique is assumed to validate the other by fitting the QSAR model into the binding domain of its target protein. These combination approaches are highly synergistic and provide information beyond the individual models (37) as exemplified below.

Combined modeling of the apical sodium-dependent bile acid transporter (ASBT)

ASBT has a critical role in lipid and cholesterol homeostasis and is a potential drug target for hypercholesterolemia as well as a drug-transporting vector for increasing bioavailability. As a result, ASBT has been studied extensively using both biochemical methods and in silico techniques. Baringhaus and co-workers generated a Catalyst pharmacophore model based on 17 chemically diverse ASBT inhibitors, which described molecular features that are essential for ASBT affinity (62). This data are in good agreement with a previous CoMFA model generated by our laboratory (52). The model also enabled the in silico discovery of substrates for ASBT. Together, the two models provided the first step in the rational design of prodrugs for specific targeting to ASBT.

To make the study more complete and visualize the transporter directly an in silico model for ASBT was generated using knowledge-based comparative modeling (5). The

25 transmembrane domains were modeled using bacteriorhodopsin as a scaffold and the extracellular loops were modeled by remote threading homology modeling (63). A subsequent docking study with the natural substrate, cholic acid, identified putative substrate binding sites that provided rational leads for site-directed mutagenesis studies.

A more recent publication from our group reinforced the theoretical model with additional biochemical data (4).

1.3 Significance of this dissertation work

Absorption into the systemic circulation is the first step for drugs to reach their targets followed by distribution to tissues. Drugs are, in many cases, then metabolized into more readily excretable forms and eliminated from the systemic circulation. All of these aspects are significantly mediated or influenced by transporters and receptors. This complex interplay of different proteins coordinates to absorb nutrients and protect against the accumulation and toxic compounds. Nucleoside transporters (chapter 2) and peptide transporters (chapter 3) facilitate absorption of a large number of drugs; organic cation transporter (chapter 4) and organic anion transporting polypeptides (chapter 5) influence drug distribution and elimination process; P-glycoprotein (chapter 6 and chapter 7) is the major drug efflux pump; androgen receptor (chapter 8) and bile acid receptor (chapter 9) represent important drug targets. Six peer-reviewed papers have been published from the above studies up to the time of writing of this dissertation.

Besides generating traditional pharmacophore models, we have also developed a novel approach called meta-pharmacophore modeling to study multiple transporters at the same time using data acquired from different laboratories and cell systems. By combining

26 pharmacophore modeling and meta-analysis, we successfully extended the application of traditional pharmacophore modeling to OATP transporters, where biological data is scarce and scattered across different laboratories and cell systems (chapter 5). This novel approach could be easily applied to other systems.

Most current pharmacophore studies are limited to model generation and interpretation. We have gone beyond these limitations by applying thoroughly validated pharmacophore models in database screening and successfully identified novel inhibitors for hPepT1 (chapter 3) and P-gp transporter (chapter 7).

By applying an array of in silico methods to a large number of pharmaceutically significant transporters and receptors in this dissertation, we have gained valuable insight into drug-transporter or drug-receptor interaction that would otherwise be unavailable using experimental methods. In addition, a comparison of the application of different in silico methods to each specific study could enable us to evaluate the advantages and disadvantages as well as the applicability of each in silico method.

Figure 1.1. General protocol for modeling of transporter and receptor proteins. A. Comparative modeling of Human ASBT 3D model (right) based on its template bacteriorhodopsin crystal structure (left). 1The template and the target transporter must have identical transmembrane topology and at least 20 % sequence identity. Although not essential, a template structure determined by X-ray crystalization is preferred. B. Putative binding or ligand-protein interaction domains can be determined through a docking study. This figure illustrates three potential high-affinity binding domains on ASBT for cholic acid, a natural susbstrate. C. General example of a 3D-QSAR modeling approach. Here, the CoMFA coefficient contour maps surrounding a model substrate for hCNT1 (upper left), hCNT2 (upper right) and hENT1 (bottom). For hCNT1 and hENT1, steric contour maps indicate that greater inhibition is correlated with less steric bulk near green contours and more steric bulk near yellow. Electrostatic contours suggest that more negative electrostatic charge near blue and more positive charge near red will increase biological activity. For hCNT2, the contours of the H-bond acceptor field map are shown in yellow and green and those of the H-bond donor field map are shown in red and blue. Greater inhibition is correlated with weaker H-bond acceptor near green, stronger H-bond acceptor near yellow, stronger H-bond donor near blue, and weaker H-bond donor near red. 2 the number of substrates (or inhibitors) ≥10; the spread of biological activity is greater than 2 orders of magnitude; all substrates are binding to the same site. D. General example of a pharmacophore modeling approach. The fit of substrate bilirubin with human Organic Anion Transporting Polypeptides C (OATPC) pharmacophore. Different spheres illustrate important chemical features resposible to OATPC binding. Cyan spheres represent hydrophobic feature and green spheres for H-bond acceptor.

Figure 1.1

Template TM Method Docked substrates Reference

ASBT Bacteriorhodopsin crystal structure 7 Knowledge based Cholic acid (4, 5) GLUT1 Helix packing model for LacY 12 Knowledge based N/A (13) LacY crystal structure 12 Not Reported N/A (15-17) Glucose, forskolin, phloretin, GlpT crystal structure 12 Nest (64), Modeller (65) cytochalasin B G6PT GlpT crystal structure 12 Modeller N/A (12) EmrE Biochemical and biophysical study results 8 (66) TPP+ (67) P-gp MsbA crystal structure (E. Coli.) 12 Modeller N/A (18) MsbA crystal structure (E. Coli.), cross linking 12 Not Reported N/A (19)

MRP1 MsbA crystal structure (V. C.) 17 Modellera N/A (22) NhaA NhaA electron density projection map 12 Knowledge based amiloride (68) DAT NhaA model generated above 12 Knowledge based S-citalopram, cocaine (60) SERT DAT model generated above 12 In silico mutation S-citalopram, cocaine (69)

30 NET SERT model generated above 12 In silico mutation S-citalopram, cocaine (69) aOnly the 12TM that are essential to function were modeled

Table 1.1. Summary for membrane transporter comparative modeling studies

Model training set Model features Correlation Program Ref P-gp 27 inhibitors of Digoxin transport (Caco-2) 4 hydrophobes, 1 HB acceptor 0.77 Catalyst (70) 21 inhibitors of vinblastine binding 1 hydrophobes, 3 ring aromatic features 0.88 Catalyst (CEM/VLB100) 17 inhibitors of vinblastine accumulation (LLC- 4 hydrophobes, 1 HB acceptor 0.86 Catalyst PK1) 18 inhibitors of calcein accumulation (LLC- 2 hydrophobes, 1 HB donor, 1 ring aromatic feature 0.76 Catalyst PK1) 16 inhibitors of verapamil binding (Caco-2) 2 hydrophobes, 1 HB acceptor, 1 ring aromatic 0.96 Catalyst (70) feature Verapamil and digoxin 5 hydrophobes, 2HB acceptor N/A Catalyst (70) 9 glucocorticoids exported by L-MDR1 4 hydrophobes, 3 HB acceptor N/A DISCO (29) 10 glucocorticoids exported by L-MDR1 Importance of nonpolar bulky group at 6-alpha 0.99 (0.48*)/ CoMFA position 0.95 (0.41*) /CoMSIA hOCT1 22 inhibitors of tetraethylammonium uptake 3 hydrophobes, 1 positive ionizable 0.86 Catalyst (31) (HeLa)

31 23 inhibitors of tetraethylammonium uptake 5 molecular descriptors correlates best with 0.95 Cerius (HeLa) inhibition

hOCT2 30 inhibitors for TEA transport hydrophobic interaction most important 0.81 Cerius 31 inhibitors for TEA transport molecular size and shape important; OCT2 binding 0.97 (0.60*) CoMFA site maybe multispecific Catalyst hCNT1 27 nucleoside analogues inhibiting thymidine 2 hydrophobes, 3 HB acceptors N/A DISCO (33) transport 35 nucleoside analogues inhibiting thymidine steric and electrostatic interaction equally important 0.98 (0.65*) CoMFA transport

Continued

Table 1.2 Summary for membrane transporter QSAR studies

Table 1.2 continued

hCNT2 13 nucleoside analogues inhibiting inosine 2 hydrophobes, 3 HB acceptors N/A DISCO transport 32 nucleoside analogues inhibiting inosine H-bonding interaction determines inhibition 0.83 (0.52) CoMFA transport hENT1 27 nucleoside analogues inhibiting uridine 2 hydrophobes, 2 HB acceptors N/A DISCO transport 39 nucleoside analogues inhibiting uridine steric and electrostatic interaction equally important 1.00 (0.74) CoMFA transport ASBT 17 bile acid reabsorption inhibitors 3 hydrophobes, 1 HB acceptor, 1 HB donor 0.94 Catalyst (62) 25 bile acid analogues tighter fit at ring system attached hydroxyl groups; 0.96 (0.63*) CoMFA (52) sterically more favorable at C17 position; electrostatic interaction mainly located at C24-C27 region DAT 50 mazindol analogues inhibition sterically favorable around heterocyclic ring A and ring 0.90 (0.70*) CoMFA (58) D; electrostatic interaction mainly around ring D 76 benztropine analogues inhibition sterically favorable in the diphenylmethoxy terminus; 0.92 (0.63*) CoMFA (59) electrostatic interaction is mainly located toward the

32 diphenylmethoxy group

SERT 12 imipramine analogues distance between fused ring and amine important for 0.53 CoMFA (71) activity 11 antidepressants Similar to above 0.55 CoMFA 10 fluoxetine analogues aromatic ring conformation determine potency 0.60 CoMFA BBB choline 33 inhibitors to choline uptake into brain positive charge; steric favorable area around alkyl 0.95 CoMFA (72) transporter chain-seqment of choline and N-methyl group; binding (0.47*)/0.85 /CoMSIA pocket is larger than choline molecule (0.37*) PEPT1 79 dipeptide-type substrates with affinity data electrostatic, lipophilic and HB donor interaction 0.90 CoMFA/CoM (35) important (0.64*)/0.91 SIA (0.78*) Gly-Sar, bestatin and enalapril 2 hydrophobes, 1 HB donor, 1 HB acceptor, and 1 N/A Catalyst (34) negative ionizable *numbers between brackets indicate cross-validated r2 values, also named q2 values; N/A: not reported.

CHAPTER 2

MOLECULAR REQUIREMENTS OF THE HUMAN NUCLEOSIDE

TRANSPORTERS HCNT1, HCNT2, AND HENT1

2.1 Introduction

Nucleoside transporters play an important role in physiology by regulating the extracellular concentration of adenosine and by salvaging nucleosides (73). Nucleoside transporters can be divided into two broad classes (74), equilibrative and concentrative.

Whereas hENT1 (es), a member of the equilibrative nucleoside transporters (ENTs), is expressed ubiquitously, Na+-dependent concentrative transporters (CNTs) are found in more specialized tissues important for absorption (intestinal epithelia), distribution

(blood-brain barrier), and elimination (hepatic and renal epithelia) of drugs. Among the nucleoside transporters, the purine specific concentrative nucleoside transporter (hCNT2), the pyrimidine specific concentrative nucleoside transporter (hCNT1), the

nitrobenzylthioinosine-sensitive (hENT1) and -insensitive (hENT2) equilibrative

transporters are expressed in the human intestinal epithelial cells (75, 76). Here they seem to be simultaneously expressed on different faces of the epithelial cells, mediating vectorial transport of nucleosides and nucleoside drugs (77).

Many antiviral (e.g., ribavirin) and anticancer nucleoside drugs (e.g., 5- fluorouridine) are substrates of nucleoside transporters (78). Their ability to transport nucleoside drugs is critical to the therapeutic effectiveness or toxicity of these drugs (78,

79). Therefore, understanding the basic molecular mechanism(s) of nucleoside transport should enable the design of more effective nucleoside drugs and those with better absorption profiles. Currently, the rational design of these drugs is hindered by the absence of high-resolution structural data on these transporters. For this reason, three- dimensional quantitative structure-activity relationships (3D-QSAR) can provide a helpful tool to direct the discovery of novel lead compounds with affinity for specific nucleoside transporters. This approach has been applied successfully to generate pharmacophore and 3D-QSAR models for the apical bile acid transporter (52), peptide transporters (80), P-glycoprotein (30), organic cation transporter 1 (31); for a recent review, see (5). Previously, a limited QSAR model was developed to map the ENT transporter binding environment (81). At that time, specific information about the presence of different subtypes of nucleoside transporters was not available. Thus, this model represents the combined activities of equilibrative nucleoside transporters from various tissues and species (HL60 human leukemia cells, human red blood cells, L1210 mouse lymphocyte leukemia cells, guinea pig myocytes). Because hCNT1, hCNT2, and hENT1 are expressed in tissues important for drug disposition, in the current study we

34 have generated distinctive pharmacophore models for each of these nucleoside transporters using the distance comparisons technique (41). We used inhibition profiles of hCNT2, hCNT1, and hENT1 transporters from our laboratory (82, 83) to gain insight into the different binding mechanisms and requirements of these three nucleoside transporters.

Subsequently, to generate 3D-QSARs, we performed comparative molecular field analysis (CoMFA) (39), which provides subtle and unique structure-activity correlations for each individual nucleoside transporter. The quality of the models was assessed by their ability to successfully predict the inhibition of a set of test compounds. The current models enable us to predict transporter affinity and guide the design of novel lead compounds for drugs that may selectively target specific nucleoside transporter isoforms.

2.2 Materials and Methods

Biological Data

The biological (nucleoside uptake) data used here have been published previously

(82, 83). Briefly, the uptake of 3H-labeled prototypic substrates of the hCNT2 (inosine,

0.5 µM), hCNT1 (thymidine, 1 µM), and hENT1 (uridine, 10 µM) transporters was measured in the presence and absence of various nucleosides and nucleoside analogs (82,

83). The nucleoside analogs (Figs. 1, 2, 3) used were mostly analogs of uridine or adenosine and had a single substitution on either the sugar ring or the base or a single substitution on both rings. In addition, several nucleoside drugs (e.g., azidothymidine, cytarabine) were also tested. These uptake data, in the presence of inhibitors (0.1 mM for hCNTs and 2 mM for hENT1) used in the QSAR analysis were expressed as the

35 percentage of the uptake obtained in the absence of these inhibitors. The hCNT1/2 studies were conducted with brush border membrane vesicles isolated from the human intestinal epithelia. Because both these transporters are expressed there, these studies could be conducted using the same vesicle preparations. We have previously shown that hCNT3 is not functionally expressed in the human intestinal epithelia (84). The hENT1 studies were conducted with the Xenopus laevis oocytes expressing recombinant hENT1 (82). In the interest of brevity, the reader is referred to publications containing these data for further information (82, 83).

Molecular Modeling and Structure Building

Details of different methods are already described in the previous chapter. Only customizations specific to each project are described here.

DISCO

Only active inhibitors were included for pharmacophore generation. The active inhibitors are defined as compounds that can introduce a statistically significant (p < 0.05) change to the transporter affinity. The reference compound was selected based on the recommendation that it should be the molecule with few features and few conformers

(41). 2'-deoxyuridine, 5'-deoxythymidine, and 5'-deoxythymidine instead of a natural substrate were selected as reference compounds for hCNT2, hCNT1, and hENT1 data sets, respectively. The feature distance tolerance was set at 1.0 Å. DISCO was initially run considering all the potential "feature" points. Additional runs with the specification of

a minimum of two hydrophobic centers were also carried out since all inhibitors contain

these two features and they make good alignment points.

FieldFit

The template molecules for hCNT1, hCNT2, and hENT1 substrates are thymidine,

inosine, and uridine, in that they represent model substrates for each data set; furthermore,

these molecules were chosen as reference compounds in collecting experimental data

(inhibition studies).

CoMFA

CoMFA model generation steps were fully described in the introduction chapter.

The test set compounds were selected from literature references (79, 85, 86). The different environment in which the experimental data for the test set and training set compounds were obtained (i.e., different expression systems, substrate concentration, and inhibitor concentrations) does not allow for a direct comparison of predicted versus experimental values. Instead, the inhibitors are categorized into active (60% transport of the reference substrate in the presence of inhibitor) and inactive (>60%) inhibitors.

Where only IC50 values are available (86) (hENT1), a compound with an IC50 2.0 mM (1

µM substrate) is considered an active inhibitor, whereas a compound with an IC50 > 2.0 mM is considered an inactive inhibitor.

2.3 Results and Discussion

To gain an understanding of the binding mode of nucleoside analogs to their transporters, we used a computational approach to model in vitro affinity data. The program DISCO was used to generate pharmacophore models of the three-dimensional orientation of essential ligand characteristics that might ultimately relate to features within nucleoside transporters. These models were derived using multiple conformations of each individual ligand alongside the experimental inhibition data. The result is a computational model that can be used to predict the affinity of nucleoside analogs to each individual transporter and serve as a guide to the design of novel transporter inhibitors and ligands.

The structures of the nucleoside analogs used in this study are shown in Figure

2.1, 2.2, 2.3, sorted by nucleobase moiety (adenine/guanine, thymine, and cytosine).

Where available, trivial and registered drug names are listed. In previous literature, the 9-

D-arabinofuranosyl conformers of adenine, cytosine, and hypoxanthine have been named

Ara-A, Ara-C, and Ara-H, respectively. The majority of DISCO feature points are located on the nucleoside pentose ring (Figure 2.4, A–C). A detailed inspection shows that besides obvious similarities among the three individual transport pharmacophores, such as two hydrophobic centers and one H-bond acceptor on the pentose ring, subtle differences set the individual transporters apart. For example, the presence of both a H- bond acceptor and donor feature near 3'-C is important for hENT1 inhibitors (Figure

2.4C); furthermore, H-bond acceptors on 3'-OH and the 2-position of the pyrimidine ring are important for both hCNT1 and hCNT2 inhibitors (Figure 2.4, A and B), whereas a H- bond acceptor on 5'-OH is important only for hCNT2 inhibitors (Figure 2.4B). In general,

38 a pentose ring structure and H-bond formation over 3'-C are necessary for a compound to effectively inhibit all three nucleoside transporters. This observation is consistent with the biological data, whereas the presence of 3'-OH is necessary for a nucleoside to be a high- affinity inhibitor or substrate of these transporters (82, 83). Furthermore, a H-bond acceptor at the 2-position of the pyrimidine ring is necessary for hCNT1 and hCNT2 inhibitors and an additional 5'-OH is required for hCNT2 inhibition. The interatomic distances between pharmacophore feature points are listed in Table 1, further illustrating similarities and distinct differences between the three transporter pharmacophore models.

In general, interatomic distances between analogous pharmacophoric points are closely correlated (e.g., hydrophobic moiety 1–2, 3.69 ± 0.17 Å; acceptor atom 2-hydrophobic group 2, 1.20 ± 0.01 Å). Despite the distinct differences between the three transporters, there are molecules that adhere to all three requirements and will therefore exert affinity for all transporter isoforms, such as uridine and adenosine.

A pharmacophore is a useful approach to analyze the chemical groups and their three-dimensional orientation required for biological activity; however, it takes into consideration only active molecules, even though inactive molecules could further delineate the chemical boundaries of designing transport inhibitors. Furthermore, pharmacophore models generally do not take into consideration potential electrostatic and steric interactions. To address these limitations, we extended our analyses using the 3D-

QSAR algorithms CoMFA and GOLPE. After excluding eventual outliers in an iterative process, predictive QSAR models were derived for each of the individual nucleoside transporters (Table 2.2). As a further illustration, C and D in Figure 2.5, 2.6, 2.7 display the residual plots of divergence between predicted and actual activity values for the

39 hCNT1, hCNT2, and hENT1 models, respectively. The CoMFA models feature robust q2 values (0.65 for hCNT1, 0.516 for hCNT2, and 0.739 for hENT1), indicative of an internally consistent model.

The biological implications of the QSAR analyses are explained through CoMFA coefficient contour maps (A, Figure 2.5, 2.6, 2.7), which illustrate the correlation of steric and electrostatic fields with biological activity. An exception is the map of hCNT2, which visualizes the H-bond aspect, in that this parameter correlates better with biological data (Figure 2.6A). A model substrate for each transporter is displayed within the contour map to facilitate interpreting the relative positioning of the CoMFA fields.

The polyhedra in each map surround all lattice points at which the QSAR strongly associates changes in interaction field values with changes in biological activity (i.e., percentage inhibition). The contours of the steric map (or H-bond acceptor field map) are shown in yellow and green, and those of the electrostatic map (or H-bond donor field map) are shown in red and blue. Greater inhibition is correlated with less bulk (weaker H- bond acceptor) near green, more bulk (stronger H-bond acceptor) near yellow, more negative charge (stronger H-bond donor) near blue, and more positive charge (weaker H- bond donor) near red.

The CoMFA plot for hCNT1 (Figure 2.5A), reveals the presence of blue contours spanning both 3'- and 5'-OH, indicating the importance of electronegative charge around these two positions and their critical role in hCNT1 transporter inhibition. An additional advantage of CoMFA over pharmacophore mapping is evident here: the 5'OH is not identified in the hCNT1 pharmacophore model because 5'-deoxyadenosine is an active inhibitor; however, the low affinity for the transporter (26.5% inhibition of reference

40 substrate) is weighed appropriately in CoMFA and correlated with the absence of a 5'-OH.

The red contour over the 3,6-position of the pyrimidine ring indicates more positive groups on these positions will contribute to stronger inhibition of hCNT1, whereas a blue contour under the 6-position of the pyrimidine ring indicates that negative groups correlate better with hCNT1 transporter inhibition. Therefore, substituting hydrogen with a hydroxyl group at the 6-position of the pyrimidine base would satisfy both requirements.

The green contour over the nucleoside base ring (Figure 2.5A), which would suggest that less bulk in this general area could result in a molecule with higher affinity to hCNT1, is predominantly produced by 8-bromoadenosine (data not shown). The base ring of 8- bromoadenosine is aligned out of the plane formed by all other compounds (not shown) and the algorithm, which can be sensitive to molecular overlap, inaccurately correlates this bulk with its low inhibitory capacity. We can therefore conclude that this polyhedron is an artifact of the molecular overlapping algorithm.

The hCNT2 CoMFA model (Figure 2.6A) differs from the hCNT1 model in that biological activity is correlated exclusively to H-bond (HB) fields compared with a combination of steric, electrostatic (CoMFA), and H-bond (GOLPE) fields for hCNT1.

Inspection of the hCNT2 field contour map reveals a predominance of H-bond donor polyhedra (blue contours), suggesting that active inhibitors of hCNT2 feature multiple H- bond donor groups. Conversely, this indicates that the binding site of the hCNT2 transporter protein may be rich in amino acids with strong H-bond acceptor features, such as Thr, Ser, Gln, or Asn. Loewen (87) showed that a highly conserved putative binding domain confers pyrimidine selectivity to hCNT1 containing Ser319/Gln320 and

Ser353/Leu354; when these amino acids were mutated to their corresponding residues in

41 hCNT2 (Gly313/Met314 and Thr347/Val348), the resulting transporter became purine- selective. Additional point mutations had variable effects on the purine and pyrimidine selectivity of both transporters. For example, the hCNT1/S319G/Q320M/S353T/L354V mutant conferred full hCNT2 transport characteristics, even though mutation of Ser319 of hCNT1 to Gly by itself enabled transport of purine nucleosides. Although the S353G mutation may imply that, relative to hCNT2, the hCNT1 binding domain comprises more

H-bond-forming amino acid residues, most mutations exchanged amino acids of similar properties. Overall, these studies do not agree or disagree with our prediction that hCNT2 inhibition is governed predominantly by amino acids capable H-bond formation. As suggested by the authors (87), some mutants may exert their effects through altered helix packing, thereby indirectly affecting substrate specificity of the transport translocation domain. Clearly, these ligand-protein interaction effects fall outside the scope of extrapolation possible with our current analysis.

The blue contour (Figure 2.6A, α field) under the 3' and 5' positions of the sugar ring of hCNT2 emphasizes the importance of the 3'-H and 5'-OH as HB donors. The green contour,χ, under the 5' and 1' positions emphasizes the role of 1'-O and 5'-OH as

HB acceptor features. The blue contour, designated β (Figure 2.6A), over 3'-OH emphasizes the importance of 3'OH in its role as a HB donor for affinity toward hCNT2.

Overall, these data emphasize the importance of the 3' hydrogen and hydroxyl as well as the 5' hydroxyl groups for effective hCNT2 transporter inhibition, which correlates well with the pharmacophore model (Figure 2.4B). From this model, we can expect a higher inhibition if more potent HB donating groups were placed around the purine base at the

2-position or possibly the 1-, 6-, or 7-position (groups R, R1, and X in Figure 2.2). For

42 example, a hydroxyl group could satisfy both HB acceptor (green) and HB donor (blue) fields around the 2-position of the purine base. These guidelines can aid in the future design of compounds that selectively inhibit hCNT2.

Analogous to hCNT1, the biological activity of hENT1 inhibitors correlates well with steric and electrostatic fields (Figure 2.7A). The large blue contour over the 3' position of the pentose moiety indicates the importance of a hydroxyl group for affinity to the hENT1 transporter. Whereas hCNT1 and hCNT2 displayed blue contours over the 5' and 2' positions, these are not present in hENT1, indicating the relative insensitivity of hENT1 transporter toward 2'- and 5'-OH. The red contour over the 2,7-position of purine

(or the 3,5-position of pyrimidine) indicates that more electropositive groups at these positions will contribute to higher affinity (i.e., inhibition) of hENT1. At the same time, the blue contour over the purine (pyrimidine) ring suggests the existence of more negative groups near the 1,2,6,7-position of purine or the 3,4,5-position of pyrimidine. A combination of these two criteria (i.e., addition of a hydroxyl group at 2,7-position of purine or 3,5-position of pyrimidine) would aid in developing more potent inhibitors of hENT1 transport. Addition of a hydroxyl group at the 7-position of purine is chemically unfeasible, but because of the inexact nature of QSAR models, addition of a hydroxyl group at the 8-position of the purine substructure might generate the same effect.

Although such compounds are not commercially available, future studies may use these findings in the design of novel compounds for hENT1. The green contours above and below the 1,7-position of purine (or 4,5-position of pyrimidine) suggest less bulky groups on these positions might lead to higher inhibition of hENT1. As noted before, however,

43 these data could be the result of overlapping artifacts and should be interpreted with caution.

The divergence of predicted from experimental data are presented in residual plots for the respective CoMFA analyses (Figure 2.5C, 2.6C, and 2.7C) and allow inspection of the internal consistency of the models. From these plots, it is apparent that the hCNT2 model has significantly more scatter than those of hCNT1 and hENT1, which is accurately reflected by their q2 and r2 values.

The GOLPE analyses assisted in further validating the biological implications of

CoMFA models. The resulting GOLPE models are statistically robust, as indicated by their q2 values (0.69 for hCNT1, 0.69 for hCNT2, and 0.70 for hENT1). GOLPE differs from CoMFA in the utilization of its probe atom, a phenolic hydroxyl group capable of donating and accepting one H-bond, allowing for an alternative interpretation of coefficient contours. The GOLPE analyses can correlate biological activity with H- bonding patterns. Because the OH probe is partially negative, its interaction with H-bond donors will be favorable (yellow contours) and its interaction with H-bond acceptor will be unfavorable (cyan contours).

The hCNT1 contour map (Figure 2.5B), reveals a yellow contour over 3'-(O)H and a cyan contour over 3'-O(H), emphasizing significance of HB donation at 3'-(O)H and HB acceptance at 3'-O(H), and further implying an important role for the dual HB donor/acceptor character of the 3'-OH in affinity for the hCNT1 transporter. This agrees with both the pharmacophore model and CoMFA result. On the other hand, GOLPE does not display a correlation with the 5'-OH group that CoMFA revealed. The small yellow contour over the 6-position of pyrimidine emphasizes the importance of H-bond donor,

44 whereas several cyan contours over the same position emphasize the importance of H- bond acceptor. A substitution of -H by -OH at the 6-position of pyrimidine would provide both H-bond donor and acceptor. This GOLPE prediction is in good agreement with suggestions from the above CoMFA analyses, further emphasizing the complementary and sometimes supplementary character of two independent 3D-QSAR techniques.

The hCNT2 GOLPE data (Figure 2.6B) show distinct yellow contours over 5'-

(O)H, 3'-(O)H, and 3'-H, indicating that both 3'-(O)H and 5'-(O)H are essential in maintaining inhibition. The cyan contours over 3'-O(H) and 5'-O emphasize the importance of 3',5'-O for hCNT2 transporter inhibition. GOLPE confirms the importance of both 3'- and 5'-OH, which were previously identified by the pharmacophore and

CoMFA models. The GOLPE contour plot, however, provides more detailed and punctate fields that can be associated with specific moieties on the substrate molecule.

The yellow contours over the nucleobase ring follow a similar pattern compared with the

HB CoMFA model. In particular, the yellow contour flanking the 2 position of purine

(favoring a H-bond donor on 2 position) confirms the previous CoMFA prediction that 2-

OH substitution of 2-H of purine could increase affinity toward hCNT2.

The yellow contours over 3'-(O)H, as well as the cyan contour over 3'-O(H), emphasize the importance of the 3' position in inhibiting the hENT1 transporter (Figure

2.7B). H-bonding fields are not detected over 2'- and 5'-OH, which is in good agreement with the pharmacophore and CoMFA models. At the nucleobase side, two yellow contours over the 3,5-position of pyrimidine (or, analogously, the 2,7-position of purine) suggest that HB forming in these two positions will increase inhibition, again confirming the above CoMFA predictions. The cyan contour near the 3-position of pyrimidine

45 provides additional confirmation of the putative positive effect of 3-hydroxyl substitution on hENT1 inhibition.

The models were subsequently used to predict the inhibition capacity of a set of test compounds (Table 4). Predictions agree with experimentally determined inhibition profiles except for 9--D-arabinofuranosyl(AF)-hypoxanthine. Contrary to literature data

(85), our model predicts this particular compound as an inactive hCNT2 inhibitor, but it correctly predicts hypoxanthine to be inactive. It is important to point out, however, that the AF analogs in our training set [i.e., compounds 9 (Fludarabine) and 30 (vidarabine)], which bear a hydroxyl group at the 2' position of the pentose moiety in the conformation, are not active hCNT2 inhibitors (88). Conversely, analogs comprising a -2'-OH moiety

(e.g., adenine) are active inhibitors of hCNT2. Intrinsically, the model can only extrapolate from the data within the training set, which signifies that the model will correlate a compound containing a -2'-OH as inactive toward hCNT2. The observation by

Lang and colleagues (85) that AF-hypoxanthine is an active inhibitor for hCNT2 could be the result of different experimental conditions. For example, both AF-adenine (31%) and adenosine (2%) were reported to be active hCNT2 inhibitors, whereas our data (83) show

AF-adenine to be inactive toward hCNT2.

Overall, steric and electrostatic interactions play an important role in hCNT1 and hENT1 inhibition, whereas H-bonding is dominant in hCNT2 inhibition. Despite their differences, all three QSAR models identified the essential role of the 3' hydroxyl group as essential for inhibiting nucleoside transporters. In addition, the 5'-OH was identified as important for hCNT1 and hCNT2 transporter inhibition. The hENT1 transporter is more tolerant to hydroxyl group substitution on the pentose ring than the other two transporters.

A recent study by Zhang and colleagues (89) on uridine-binding motifs of hCNT1 and hCNT3 concluded that the following groups determined substrate affinity for hCNT1 in decreasing sensitivity: C(3') > C(5') = N(3) > C(2'). This study is in good agreement with the DISCO-generated pharmacophore points for hCNT1 (i.e., both 3'- and 5'-hydroxyl groups). The conclusion that pyrimidine N(3) is involved in H-bonding does not concur with our results that its neighboring C(2)=O forms a H-bond acceptor feature. However, the frequently observed tautomeric forms of the pyrimidine base caused by a highly delocalized N(3) hydrogen atom may explain this minor discrepancy.

Both CoMFA and GOLPE models suggest that addition of a hydroxyl group at the 6-position of pyrimidine analogs would increase the inhibition of hCNT1; addition of a hydroxyl group at the 2-position of purine analogs would increase the inhibition of hCNT2; and addition of a hydroxyl group around the 3,5-position of pyrimidine (or 2,8- position of purine) analogs would increase the inhibition of hENT1.

A comparison of pharmacophoric features among the three models demonstrates that hENT1 is least sensitive to inhibitor modifications whereas hCNT2 is the most sensitive transporter, which is in agreement with experimental data. At first glance, it may seem extraordinary that H-bond formation nearby 3' carbon is a required feature for hENT1 affinity, even though a well characterized substrate such as 2'3'-dideoxyadenosine does not contain a carbonyl or hydroxyl group at this specific location. Upon further inspection of the DISCO features on 2',3'-dideoxyadenosine (Figure 2.4D), however, it becomes apparent that the H-bonding feature over 3'-C is actually derived from the lone pair at 5'-oxygen, whereas the H-bond acceptor feature derives from the 3-amine of the base. Thus, based on experimental results alone, H-bonding over 3'-C seems not to be an

47 essential requirement for hENT1 inhibition, but our study shows that alternate features near this group may compensate for the lack of this moiety on the base. The fact that

DISCO analysis discloses an alternate pathway for H-bond formation near 3'-C demonstrates the ability of DISCO to discriminate alternative bioactive conformations of transport inhibitors. Thus, this analysis reveals the possibility of designing novel nucleoside analogs that retain affinity for hENT1 despite the absence of a requisite 3'-OH moiety.

A recent paper by Huang and colleagues (90) describes the effective inhibition of nucleoside transport by the p38 MAPK inhibitors SB202474, SB203580, and SB203580- iodo (Figure 2.8A). A structural analog of these compounds, SB220025, reportedly did not bear any effect on [3H]uridine transport in K562 cells, which are known to express the equilibrative nucleoside transporter hENT1. These non-nucleoside analogs were subsequently built, energy-optimized, and aligned to uridine by overlapping the pentose ring to the imidazole moiety of all test structures. The alignment of the uridine ribose ring to the methylsulfoxide-phenyl (or methoxyphenyl) ring resulted from a field-fit energy minimization. Both imidazole and methylsulfoxide-phenyl moieties satisfy the hydrophobic features in hENT1 pharmacophore (Figure 2.8B). Correspondingly, the imidazole 1-nitrogen in the MAPK inhibitors represents the H-bond acceptor pharmacophore feature on the pentose ring of uridine. Based on pharmacophore features alone, the inactivity of SB220025 toward hENT1 can be explained by the substitution of the methylsulfoxide-phenyl group with a piperidine moiety on the imidazole nitrogen

(Figure 2.8A). In the active molecules, the fluoride (or iodo) atom and the highly electron deficient hydrogen atoms flanking the halide atom could act, respectively, as H-bond

48 acceptor and donor moieties, thus representing the corresponding pharmacophore groups over the 3-pentose position of uridine.

CoMFA predictions for hENT1 inhibition of SB203580, SB203580-Iodo,

SB202474 and SB220025 were 57.7, 61.8, 64.4, and 68.6%, respectively. It is interesting to note that, even though the pharmacophore model effectively excluded SB220025,

CoMFA-based predictions are unable to differentiate between the p38 MAPK inhibitors.

Upon inspection of the CoMFA fields, however, it became apparent that the model is expectedly silent in areas where the training set lacks structural diversity (Figure 2.8C).

The piperidine ring of SB220025 protrudes into such a "silent" area of the model and, consequently, does not contribute negatively to its CoMFA-based prediction of hENT1 inhibition.

None of the MAPK inhibitors are predicted to be effective inhibitors of hCNT1 or hCNT2, based on both pharmacophore and CoMFA analyses (data not shown). Future studies should be aimed at validating these predictions.

2.4 Conclusion

The structural features that are essential for nucleoside transporter affinity, transport, and inhibition have received increased attention from both a drug discovery and drug delivery standpoint. In the absence of a high-resolution protein structure, it is currently not feasible to directly design compounds that would either use or inhibit these transporter systems. In the study presented here, we examined the structural features of nucleoside transporter inhibitors and correlated these to their biological activity for each respective nucleoside transporter isoform. The resulting data present a roadmap toward

49 recognizing the molecular characteristics that are required for inhibiting each individual transporter, their commonalities as well as the features that set them apart. For example, the models can distinguish effectively between inhibitors for different isoforms; e.g., 2- deoxycytidine was predicted to be an inhibitor for hENT1, but not hCNT2.

It should be pointed out that the current models are derived from inhibition data and thus that the models predominantly represent features important for nucleoside transporter inhibition. Although care should be taken in extrapolating these findings to predict nucleoside transporter substrates, invariably many of the active inhibitors in the current study have been shown to be substrates for the transporters as well. Therefore, it is not unreasonable to speculate that active inhibitors (i.e., >60% inhibition) may very well turn out to be good nucleoside transporter substrates, when validated in vitro.

The three-pronged approach of pharmacophore mapping and two independent multivariate 3D-QSAR methods represents an indirect examination of the nucleoside transporters' binding sites. From these models, we have been able to predict the inhibition of some compounds with the individual transporters, such as p38 MAPK inhibitors.

Additional studies will indubitably improve the models and may ultimately diminish the intrinsic occurrence of false positives and negatives in each model (e.g., AF- hypoxanthine). This will require an iterative process of additional data input and evaluation of novel test compounds. In addition, the possibility of database mining in combination with de novo ligand design can be used to find lead compounds that may have affinity for nucleoside transporters or design novel chemical entities with these properties. In this respect, the use of a CoMFA-derived binding site cavity in combination with structure generating and docking algorithms, such as LeapFrog (Sybyl,

Tripos Associates, Inc), will be an effective approach. For a comprehensive review of recent de novo design approaches, please refer to (91). Our future studies are aimed at further exploration of the binding site requisites of the respective nucleoside transporters to facilitate and optimize rational drug design.

hCNT1 Hydr2 AA1 AA2 AA3 Hydr1 3.79a 2.59 3.60 6.02 Hydr2 3.94 1.20 2.56 AA1 4 5.97 AA2 3.58 hCNT2 Hydr2 AA1 AA2 AA3 AA4 Hydr1 3.49 2.95 3.52 5.54 2.75 Hydr2 3.17 1.19 2.54 2.88 AA1 3.97 4.32 3.65 AA2 3.50 2.99 AA3 4.14 hENT1 Hydr2 AA AS DS Hydr1 3.79 3.60 7.00 7.21 Hydr2 1.20 4.75 4.86 AA 5.90 6.01 DS 0.26 aAll distances in Å; AA, acceptor atom; Hydr, hydrophobic atom; AS, acceptor site; DS, donor site.

Table 2.1. Intramolecular atomic distances between pharmacophoric feature points. See Figure 2.4, A-C, for details on relative positioning of pharmacophore features on representative ligands.

Type q2 "press" r2 s #c %Steric %Electrostatic CoMFA hCNT1 0.650 25.912 0.981 6.042 5 0.46 0.54 hCNT2 0.516 24.33 0.827 14.52 2 0.51a 0.49b hENT1 0.739 21.495 0.998 2.097 8 0.47 0.53 GOLPE hCNT1 0.694 20.20 0.892 12.00 3 - - hCNT2 0.685 18.42 0.955 6.94 5 - - hENT1 0.700 20.11 0.922 10.23 4 - - a%HB Acceptor for hCNT2 b%HB Donor for hCNT2 cNumber of components used for non-cross-validated analyses

Table 2.2. Statistical parameters for 3D-QSAR analyses

Predicted (%inhibition) Actual (%inhibition) 0.1mM inhibitor, 1µM hCNT1 [3H]Inosine 0.1mM inhibitor_10µM [3H]Uridine 34 Guanosine Inactive (114.0) Inactive (~90) 45 Gemcitabine Active (41.6) Active (14) 46 2'-deoxycytidine Active (53.1) Active (~20)

0.1mM inhibitor, 0.5µM hCNT2 [3H]Thymidine 1mM inhibitor, 10µM [3H]Uridine 46 2'-deoxycytidine Inactive (95.1) Inactive (96) 47 2'-deoxyguanosine Active (44.7) Active (3) 48 2'-deoxyinosine Active (52.7) Active (5) 49 5-methyuridine Active (23.2) Active (38) 50 AF-hypoxanthine Inactive (106.7) Active (20) 51 Hypoxanthine Inactive (76.6) Inactive (96)

2mM inhibitor, 10µM hENT1 [3H]Uridine IC50 mM, 1µM [3H]Uridine 46 2'-deoxycytidine Active (27.6) Active (2.0) 48 5-methyluridine Active (21.2) Active (0.17) 20 5-fluoro-5'-deoxyuridine Active (29.6) Active (0.15) 52 5-bromo-2'-deoxyuridine Active (41.7) Active (0.24) 53 3'-deoxycytidine Active (56.3) Active (3.1) AF=9β-D-arabinofuranosyl

Table 2.3. QSAR predictions of test compounds (CoMFA)

Figure 2.1. Chemical structures of uridine analogs and drugs used in QSAR analyses. The adenine base and deoxyribonucleoside groups are numbered for reference.

Figure 2.2. Chemical structures of adenosine analogs and drugs used in QSAR analyses. The adenine/guanine base and deoxyribonucleoside groups are numbered for reference.

Figure 2.3. Chemical structures of cytidine analogs and drugs used in QSAR analyses. The cytosine and deoxyribonucleoside groups are numbered for reference.

Figure 2.4. Inhibition pharmacophore models for the three nucleoside transporters in relationship to the reference compound for each data set. A, hCNT1 is visualized in reference to its active inhibitor 5'-deoxythymidine; B, hCNT2 pharmacophore features are shown on 2'-deoxyuridine; and C, hENT1 pharmacophore moieties are presented on its substrate 5'-deoxythymidine. D, DISCO features of the hENT1 substrate and inhibitor 2',3'-dideoxyadenosine, illustrating the H-bond donor and acceptor site features of 3-N and 5'-OH, respectively, in the absence of a 3'-hydroxyl moiety. DISCO features are labeled and represented by red spheres.

Figure 2.5. 3D-QSAR model of nucleoside transporter hCNT1. A, CoMFA coefficient contour map surrounding model substrate thymidine illustrating the correlation of CoMFA fields with biological activity. Steric contour maps indicate greater inhibition is correlated with less steric bulk near green contours and more steric bulk near yellow. Electrostatic contours suggest that more negative electrostatic charge near blue and more positive charge near red will increase biological activity. B, GOLPE coefficient contour map illustrating the correlation of H-bond acceptor (cyan) and donor (yellow) fields with percent inhibition. C, CoMFA residual plot indicating correlation and internal consistency of the model. D, GOLPE residual plot.

Figure 2.6. 3D-QSAR model of the nucleoside transporter hCNT2. A, CoMFA coefficient contour map surrounding model substrate inosine illustrating the correlation of CoMFA fields with biological activity. The polyhedra in each map surround all lattice points where the QSAR strongly associates changes in interaction field values with changes in biological activity (i.e., percentage inhibition). The contours of the H-bond acceptor field map are shown in yellow and green and those of the H-bond donor field map are shown in red and blue. Greater inhibition is correlated with weaker H-bond acceptor near green, stronger H-bond acceptor near yellow, stronger H-bond donor near blue, and weaker H-bond donor near red. B, GOLPE coefficient contour map illustrating the correlation of H-bond acceptor (cyan) and donor (yellow) fields with percentage inhibition. C, CoMFA residual plot indicating correlation and internal consistency of the model. D, GOLPE residual plot.

Figure 2.7. 3D-QSAR model of the nucleoside transporter hENT1. A, CoMFA coefficient contour map around the model compound uridine illustrating the correlation of CoMFA fields with percentage inhibition. Steric contour maps indicate that greater inhibition is correlated with less steric bulk near green contours; more steric bulk near yellow. Electrostatic contours suggest that more negative electrostatic charge near blue, and more positive charge near red will increase biological activity. B, GOLPE coefficient contour map illustrating the correlation of H-bond acceptor (cyan) and donor (yellow) fields with percent inhibition. C, CoMFA residual plot indicating correlation and internal consistency of the model. D, GOLPE residual plot.

Figure 2.8. Predictions for p38 MAPK inhibitors. A, structural formulas of p38 MAPK inhibitors; top, SB220025; SB202474, R1 = OCH3, R2 = CH2-CH3; SB203580, R1 = SOCH3, R2 = p-F-Ø; SB203580-iodo, R1 = SOCH3, R2 = m-I-Ø. B, alignment of p38 MAPK inhibitors onto uridine together with the hENT1 pharmacophore points. C, relative position of the p38 MAPK inhibitor structures within the hENT1 CoMFA contours.

CHAPTER 3

IN VITRO AND PHARMACOPHORE BASED DISCOVERY OF NOVEL HPEPT1

INHIBITORS

3.1 Introduction

The human intestinal small peptide carrier (hPepT1) is a proton-coupled, low- affinity, high-capacity oligopeptide transport system with broad substrate specificity. In addition to transporting its natural substrates, di- and tri-peptides occurring in food products (35), it shows affinity toward a broad range of peptide-like pharmaceutically relevant compounds, such as β-lactam antibiotics (92) and angiotensin converting enzyme (ACE)-inhibitors (93). For this reason, hPepT1 has been recognized as an important intermediate in the oral bioavailability of peptidomimetic compounds (94).

However, the lack of knowledge regarding structural specificity towards its substrates has prevented the use of this transporter on a more rational basis. Recently, Zhang and colleagues reported 9 distinct single nucleotide polymorphisms for hPepT1; only one

63 displayed a reduced transport capacity (95), inferring that hPepT1 may be a pharmacologically relevant drug delivery target not confounded by genetic variability.

Thus, there exists a keen interest in understanding the structural determinants for substrates and inhibitors of hPepT1.

Despite the availability of several computational models for peptide transporters from various species as well as hPepT1, little progress has been made in the design and elucidation of novel substrates for this key intestinal transport protein. In fact, discovery of most hPepT1 substrates has remained remarkably and perhaps unacceptably serendipitous to date. In this current study, we present a novel approach using pharmacophore-based database searching for rapidly retrieving hPepT1 inhibitors and substrates for this transporter. This method complements the many experimental approaches that are currently in use for identifying high affinity hPepT1 ligands as well as understanding the structural features for binding (96). This proof of principle study may be applicable to other human transporters for which there is significantly less data than for hPepT1.

3.2 Methods

The computational studies were carried out in collaboration with Dr. Sean Ekins.

hPEPT pharmacophore development

Catalyst version 4.8 (Accelrys, San Diego, CA) was used to generate a common features (HIPHOP) (43) pharmacophore for the well-known PepT1 substrates Gly-Sar, bestatin and enalapril (97). Structures for these 3 molecules were initially sketched in

ChemDraw version. 7.0.1 (CambridgeSoft, Cambridge, MA), exported into mol file format and then imported into Catalyst™. The three hPepT1 substrate molecules were then aligned using hydrophobic, H-bond acceptor, H-bond donor and negative ionizable features in the HIPHOP algorithm within Catalyst™.

Pharmacophore-based database searching

The resulting hPepT1 HIPHOP pharmacophore was used for a fast flexible search of the Comprehensive Medicinal Chemistry (CMC) “drug-like” database (MDL, San

Leandro, CA) of over 8000 molecules. These molecules were implemented as a Catalyst searchable database by generating up to 100 conformers with the fast conformer generation method, allowing a maximum energy of 20kcal/mol. The pharmacophore was used with the fast-flexible search approach and retrieved 145 virtual hits. This list was then sorted by molecular weight, and the top 7 molecules were individually fast-fit to the hPepT1 pharmacophore. As a means of comparison, the relatively high affinity substrates enalapril and bestatin were scored with the same method (3.74 and 2.59 arbitrary units, respectively). A search of the literature was carried out to ascertain the commercial

65 availability of these molecules for purchase. A larger range of known hPepT1 substrates or inhibitors not included in the pharmacophore but retrieved upon database searching were also scored using the same approach described above and included: Ampicillin

(2.69 arbitrary units), Captopril (2.07), Cefaclor (3.24), Cefadroxil (3.14), Cefoperazone

(2.60), Cefoxitin (2.89), Cephalexin (3.15), Cephradine (2.90), Methyldopa (2.50) and

Valacyclovir (2.13).

The pharmacophore model was also used to search an in-house database (SCUT) of 576 known drugs in clinical use in the USA derived from the Clinician’s Pocket Drug

Reference (98) in order to identify known peptide transporter substrates and inhibitors which fulfill the pharmacophore requirements. This database was created using structures in the sdf format prior to conversion to a 3D Catalyst database after generating up to 100 molecule conformations with the fast conformer generation method, allowing a maximum energy of 20 kcal/mol. The pharmacophore was then used with the fast- flexible or best search algorithms (Accelrys, Catalyst Tutorials 2003).

Fast Fit means finding the optimum fit among the existing conformers of the molecule without performing an energy minimization. Best Fit means that the conformers selected are manipulated to minimize the distances between tethered objects in the molecules, while keeping the resulting conformer energy within the Energy Limit.

Chemicals

All chemicals purchased were of the highest commercial purity. Repaglinide was purchased from TRC Biomedical Research (Toronto, Canada) and fluvastatin was

obtained from Calbiochem (La Jolla, CA). Aspartame was from Sigma-Aldrich (St. Louis,

MO). 3H-GlySar (4Ci/mmol) was purchased from Moravek Chemicals (Brea, CA).

In vitro hPepT1 bioassay

The experimental studies were performed by Dr. Jeffrey Johnston and Praveen

Bahadduri.

A stably transfected cell line (CHO-hPepT1; kind gift from Dr. Wolfgang Sadée,

Ohio State University, Columbus, OH) (13) was maintained in DMEM containing 100

µg/ml penicillin/streptomycin and 0.1% gentamycin to maintain selection of hPepT1.

Since penicillin is a potential hPepT1 substrate and its presence may interfere with

transport experiments, before each experiment cells were seeded in 24-well plates at a

concentration of 75,000 cells/well and allowed to grow in antibiotic free DMEM

supplemented with 9% FBS for approximately 48 h. Culture media was removed and

replaced with uptake buffer (3mM Hepes/Mes/Tris (pH 6.0) containing 100 mM NaCl, 2

mM KCl,, 1 mM MgCl2, and 1 mM CaCl2) for 10 minutes. After the equilibration period,

buffer was removed and replaced with fresh buffer that contained 3H-Gly-Sar (5 µM, 1

µCi/ml, S.A. 2 Ci/mM) and various concentrations of the substrate of interest (e.g. 0.033,

0.1, 0.33, 1.0, 3.3, 10, 33, and 100 mM, depending on compound solubility). After incubation with tracer and substrate for 30 min the cells were rinsed several times with ice cold buffer and lysed with Triton-X 100. Aliquots were then taken for liquid scintillation counting. The IC50 values (the concentration which resulted in 50% uptake as compared to control) were determined using a sigmoidal dose-response equation with

variable slope in GraphPad Prism version 4.0a (GraphPad, Inc., San Diego, CA). Gly-Sar

was used as an internal control for hPepT1 activity (Mean IC50 0.80 ± 0.07 mM).

3.3 Results and Discussion

This work evaluates the capability of commercially available rapid computational approaches to identify ligands for transporters as well as the validation of this method in vitro. In the current study we have used the molecular structures of well known, relatively high affinity hPepT1 substrates Gly-Sar, bestatin and enalapril to generate a HIPHOP common feature 3D-pharmacophore (Figure 3.1). These molecules represent three different classes of drugs: a dipeptide, a peptidomimetic and an ‘angiotensin converting enzyme’ (ACE)-inhibitor, respectively. The pharmacophore alignment consisted of two hydrophobic features, a H-bond donor, H-bond acceptor, and a negative ionizable feature which would indicate these are important features for interaction with this transporter in vitro.

Previous studies (35, 80, 99) have modeled peptide transport using conformational analyses to determine the molecular determinants and the distances between functional groups in substrates critical for affinity. Using relatively rigid β- lactam molecules we previously suggested that a carboxylic carbon (likely to position in a positively charged pocket), 2 carbonyl oxygen atoms (H-bond acceptors), a hydrophobic site and finally an amine nitrogen atom (H-bonding region) were important features of substrates (80, 99). Other studies using expressed rabbit PepT1 or human

PepT1 indicate a peptide bond is not essential for substrates of these ortholog transport carriers. Instead, two ionized amino or carboxyl groups with at least 4 carbon units

between them, or amino acid esters of nucleosides, such as 5-amino-pentanoic acid and

valacyclovir, permit transport (100-102). A more recent meta-analysis of Ki values for

42 substrates using PepT1 data from many sources provided a template consisting of an

N-terminal NH3 site, a H-bond to the carbonyl group of the first peptide bond, a

hydrophobic pocket and a carboxylate binding site (20). Other groups have confirmed

these essential features for PepT1 transport (103, 104). A recent article (5) described

comparative molecular field analysis (CoMFA) and a comparative molecular similarity

indices analysis (CoMSIA) models for hPepT1 using a series of 79 dipeptide-type

substrates and test set set of 19 dipeptides and dipeptide derivatives with acceptable

model statistics. CoMSIA contour maps enabled the identification of the key elements for

the binding of PEPT1 substrates. This model possibly provides a means to

computationally predict the binding of other potential hPepT1 inhibitors in the future,

however due to the structurally homologous training set and the expression of the

transporter in Caco-2 cells (which also expresses other transporters) there may be some

difficulty in extrapolating to more structurally diverse compounds.

To identify drugs that are also novel hPepT1 inhibitors we have used our

pharmacophore to search the CMC database of over 8,000 “drug-like” molecules. Prior to

screening the molecules were first converted into a multi-conformer three-dimensional

database. This allowed us to take into consideration molecular flexibility, thereby

ensuring that fast-fitting would not be limited to rigid molecules with conformations

already aligned to the pharmacophore. We retrieved 145 (~1.8% of the total database)

virtual hits mapping to the pharmacophore features. After fast fitting in Catalyst, the 7

(0.09% of the total database) best scoring molecules with drug-like molecular weight

(<500) were selected (Table 3.1) for purchase and in vitro testing. Three of these molecules (fluvastatin, aspartame and repaglinide) were readily available and mapped well to the pharmacophore features (Figure 3.2). These molecules were also experimentally identified as new inhibitors with affinity for hPepT1 that is in a range similar to Gly-Sar (≤ 1mM (8), Table 3.1). A set of 10 hPepT1 substrates or inhibitors when fast fit to the hPepT1 pharmacophore provide fit values between 2.07 and 3.24 arbitrary units. Repaglinide and aspartame possess fit values within this range (Table 3.1) whilst fluvastatin is beyond the upper end of this range. As the known hPepT1 substrates

Valacyclovir (2.13), Enalapril (3.74) and Bestatin (2.59) all have relatively similar fit values to the molecules we have tested it is likely they may behave similarly as substrates.

Interestingly, human oral bioavailability of fluvastatin is variable and low (29 ± 18%)

(23), whereas repaglinide has a more pronounced systemic availability after oral administration (56 ± 7%) (105). Aspartame (N-L-alpha-aspartyl-L-phenylalanine, 1- methyl ester) bioavailability cannot be readily assessed, but the appearance of phenylalanine and aspartate in plasma suggests that the fraction absorbed from the gut is close to unity (106). The role of PepT1 in the oral bioavailability of these three compounds had not been previously noted and this may explain to some extent the relatively high bioavailability and/or the variable pharmacokinetic and pharmacodynamic profile for these compounds. Naturally, potential drug-drug interactions with other hPepT1 substrates/inhibitors can be envisioned. It may be interesting to speculate that the widespread use of aspartame-based sugar substitute in food products and pharmaceuticals may potentially interfere with hPepT1-mediated drug absorption. Although this possibility cannot be ruled out, the relatively high IC50 of aspartame combined with its

mg dosing and rapid intestinal metabolism into phenylalanine, aspartate and methanol

would minimize the potential impact on drug absorption via hPepT1. After completion of

this study nateglinide, a molecule in the same therapeutic class and structurally similar to

repaglinide was shown to be an inhibitor but not a substrate of the rat PepT1 (107). In the present study we have identified 3 molecules as inhibitors of hPepT1 and in order to test our hypothesis that these molecules are also substrates further experiments are required. Based on a comparison of the predicted scores with the pharmacophore these molecules are in the range of the known substrates used to build the model.

We also performed a computational validation of our hPepT1 pharmacophore by searching a database of over 500 commonly prescribed drugs which we created based on a clinicians reference (98). We were able to select 65 molecules using the pharmacophore, of these 27 were known PepT1 substrates or inhibitors based on literature or our own studies. Obviously some of the other molecules which scored well using either best or fast fit algorithms may be considered for future testing.

The use of pharmacophore-based database searching with hPepT1 has demonstrated that this approach is ideal for identifying the role of peptide transport in absorption of new chemical entities prior to in vitro testing of these compounds. The method used is amenable to searching large databases of molecules with 3D conformers and has been previously applied to discovering active molecules for therapeutic targets of commercial interest, lead selection, as well as understanding the key features on P- glycoprotein substrates (108). In turn this method may also be useful for determining the role of other transporters by quickly searching for new ligands when little experimental data is available in the literature. By focusing on the key features for binding to the

71 transporter we have been able to find molecules that are either as active or more active for hPepT1 than previously known compounds used in model building. Obviously we are now in a position to iteratively improve the model using these newly discovered hPepT1 inhibitors and ultimately optimize the selectivity and specificity for future ligand identification.

Molecule Catalyst fast-fit score* In vitro IC50 for CHO-PepT1 (mM) Repaglinide 3.19 0.18 ± 0.01 Aspartame 2.39 7.69 ± 0.09 Fluvastatin 3.44 0.34 ± 0.04 Bumetanide 3.54 NT Netobimin 1.26 NT Pravastatin 3.39 NT Cerivastatin 1.27 NT *a higher score indicates a better fit to the pharmacophore

Table 3.1. Catalyst CMC database search results for selected molecules after fast flexible searching and the in vitro IC50 data for inhibition of Gly-Sar uptake in CHO-PepT1 cells . Mean ± SD of 3 experiments, NT = not tested.

Molecules Best Fit Score Fast Fit Score Molecular Weight fluvastatin 7.68 3.44 411.47 argatroban 4.83 3.57 508.64 bacitracin 4.70 2.82 1422.71 cefdinir 4.61 3.40 395.41 montelukast sodium 4.61 3.41 608.17 ceftizoxime 4.57 2.95 383.40 enalapril 4.57 3.74 376.45 lisinopril 4.56 3.18 405.49 cefixime 4.52 3.73 453.44 pravastatin 4.52 3.39 424.53 cefonicid 4.47 3.58 542.56 mupirocin 4.41 3.31 500.63 glucagon 4.38 2.92 3482.76 mezlocillin 4.33 2.99 539.58 epoprostenol 4.30 3.23 352.47 repaglinide 4.30 3.19 452.59 cefditoren 4.28 3.15 506.57 piperacillin 4.27 3.11 517.56 amoxicillin 4.21 3.13 365.40 moexipril 4.20 3.57 498.58 ceftriaxone 4.18 3.01 554.57 quinapril 4.18 3.27 438.52 ramipril 4.18 3.25 416.52 octreotide 4.17 2.87 1033.23 alprostadil 4.02 3.32 354.49 bumetanide 4.01 3.54 364.42 loracarbef 4.00 3.62 349.77 aztreonam 3.99 3.43 435.43 ceftazidime 3.96 3.29 546.57 cefmetazole 3.95 3.41 471.52 cefotaxime 3.95 3.37 455.46 ceftibuten* 3.95 3.07 410.42 dinoprostone 3.95 3.55 352.47 cefpodoxime 3.94 3.25 427.45 losartan 3.94 3.50 422.92 * cis conformation, bioactive

Continued

Table 3.2. Hit list from hPepT1 model based SCUT database search. Search results for selected molecules after Fast flexible and Best searching. Molecules in bold are known peptide transporter substrates and or inhibitors either identified in the literature or the present study.

Table 3.2 continued

ticarcillin 3.94 3.18 384.42 cefuroxime 3.93 2.97 424.39 fexofenadine 3.93 3.15 501.67 benazepril 3.91 3.37 424.50 nafcillin 3.89 3.43 414.48 perindopril 3.89 3.36 368.47 trandolapril 3.88 3.06 430.54 cromolyn 3.86 3.27 468.37 vancomycin 3.86 3.44 1449.27 amphotericin b 3.85 3.32 924.09 bivalirudin 3.84 3.19 2180.30 furosemide 3.84 2.77 330.74 tirofiban 3.79 3.24 440.60 cefotetan 3.77 2.99 575.60 cefazolin 3.73 3.52 454.50 calcitonin, human 3.66 3.38 3417.85 eptifibatide 3.66 3.28 814.93 meropenem 3.52 3.20 383.46 liothyronine 3.29 2.89 650.98 cefaclor 3.24 3.24 367.81 cephalexin 3.15 3.15 347.39 cefadroxil 3.14 3.14 363.39 levothyroxine 3.08 2.88 776.87 etodolac 3.00 2.97 287.36 nateglinide 3.00 2.93 317.43 cefoxitin 2.89 2.89 427.45 cephradine 2.89 2.90 349.40 ampicillin 2.77 2.69 349.40 captopril 2.67 2.07 217.28 cefoperazone 2.57 2.60 645.66

Figure 3.1. HIPHOP Pharmacophore for hPepT1 substrates (red = Gly-Sar; green – bestatin; yellow = enalapril). Pharmacophore features: cyan = hydrophobe; green = HBA; purple = HBD; blue = negative ionisable.

Figure 3.2. Visualization of high scoring molecules discovered with and fitted to the hPepT1 HIPHOP pharmacophore, a: aspartame, b: fluvastatin, c: repaglinide. Pharmacophore features: cyan = hydrophobe; green = HBA; purple = HBD; blue = negative ionisable.

CHAPTER 4

MOLECULAR DETERMINANTS OF SUBSTRATE/INHIBITOR BINDING TO THE

HUMAN AND RABBIT RENAL ORGANIC CATION TRANSPORTERS

HOCT2 AND RBOCT2

4.1 Introduction

The kidney plays a key role in the secretion and subsequent elimination of drugs, toxins, and other xenobiotics from the body(109-112). Many of these compounds are organic cations in that they carry a net positive charge at physiological pH, including compounds from a broad array of chemical and clinical classes (e.g., antiarrhythmics, β- adrenoreceptor blocking agents, antihistamines, antivirals, and skeletal muscle relaxing agents). Organic cations (and bases; collectively, ‘OCs’) are actively secreted by the proximal tubule by means of a two-step process (112). The first step involves transport of OC from the blood, across the basolateral membrane, into the proximal tubule cell via electrogenic, facilitated diffusion. The second step appears to be dominated by an organic cation/proton (OC/H+) exchanger located in the apical membrane that transports

78 the OC out of proximal cells into the tubular filtrate. Several of the organic cation transporters (OCTs) thought to play a role in the transport of these compounds across the basolateral membrane have been cloned in recent years, including OCT1, OCT2, and

OCT3 (110, 112, 113). OCT1 and OCT2 appear to play the predominant role in secretion of the so-called ‘Type I’ OCs (i.e., generally monovalent, hydrophilic, MW < 400; (114)) in rodent and rabbit proximal tubules (89, 115-117). Indeed, active secretion of the prototypic Type I substrate, tetraethylammonium (TEA) is eliminated in the OCT1/2 null mouse (118). In the human, however, comparatively low expression of the mRNAs for

OCT1 and OCT3, relative to OCT2 (119), supports the conclusion that OCT2 is probably the principal basolateral route for OC uptake into human proximal tubule cells. An understanding of the physical and structural characteristics that influence the binding of substrates to OCT2 would, therefore, assist in the development of models of substrate interaction with OCT2. Such models offer the promise of predicting clinically deleterious drug interactions and aiding in the design of novel pharmacological agents alongside the widely described in vitro and in vivo models used for generating much of the data on these transporters (111).

Previous studies have shown hydrophobicity and basicity to be important determinants of substrate specificity for OCTs in the apical and basolateral membranes of the rat proximal tubule (120, 121). More recent work showed that placement of planar hydrophobic mass, relative to a positively charged nitronium nucleus, is important for substrate binding to hOCT1 (31) and to the OC/H+ exchanger in rabbit renal brush border membrane vesicles (122). In contrast, relatively little is known of the structural requirements for OCT2 substrate binding, other than the critical role of the degree of

79 ionization identified by the increased IC50 values of weak bases when external pH is shifted from 7 to 8 (thereby decreasing protonation; (123)). In the present study, substrate/transporter interactions were investigated on a much larger scale in an attempt to identify distinct molecular characteristics that play a role in selectivity of OCT2. To accomplish this, a set of structurally diverse compounds was chosen for inhibition studies with the human and rabbit orthologs of OCT2 that were stably transfected in CHO cells.

This effort to develop a strategy for predicting molecular criteria that influence binding to

OCT2 involved use of two different computational methods to generate 2D- and 3D- quantitative structure activity relationship (QSAR) models. Whereas both approaches proved more effective for predicting substrate-transporter interactions than simple substrate hydrophobicity, the 3D-QSAR proved to have greater predictive power, suggesting that steric factors play a more important role in the binding process than previously acknowledged.

4.2 Materials and Methods

The experimental studies were carried out by Dr. Steve Wright and colleagues.

Chemicals

[14C]TEA (55.6 Ci /mmol) was acquired from Wizard Laboratories, Inc. (West

Sacramento, CA). 1-Methyl-4-phenylpyridinium (MPP) was from Research

Biochemicals International (RBI, Natick, MA). 1,3,5-trimethyl-4-phenylpyridinium, 2,4-

dimethyl-9-H-indenol[2,1-c]pyridinium, and 3,5-dimethyl-4-phenylpyridinium oxide

were synthesized by the Synthesis Core of the Southwest Environmental Health Sciences

Center and the Department of Chemistry at the University of Arizona (Tucson, AZ).

NBD-TMA was synthesized as described previously (124). The set of N-1 substituted

pyridiniums and quinoliniums and ethyl acridinium were synthesized as described

previously (125). All other chemicals were acquired from Sigma – Aldrich Chemicals

(St. Louis, MO) or other standard sources and were the highest grade available.

CHO cell culture and stable expression of hOCT2 and rbOCT2

Chinese hamster ovary (CHO) cells were acquired from the American Type

Culture Collection (ATCC, Manassas, Virginia) and grown in Ham’s F12 Kaighn’s

Modification (Sigma, St. Louis, MO) containing 10% fetal bovine serum (Hyclone,

Logan, UT) and maintained in a humidified atmosphere with 5% CO2. For stable expression of hOCT2 and rbOCT2, cells were electroporated with 10 µg pcDNA3.1 plasmid DNA containing the hOCT2 or rbOCT2 construct, and 10 µg salmon sperm

(Gibco, Rockville, MD) in a cuvette (4 mm gap) using a BTX ECM 630 electroporator with settings of 1050 µF, 260 V, and no resistance. Forty-eight hours following the electroporation, positively transfected cells were identified (based upon their ability to accumulate the fluorescent OC, [2-(4-nitro-2,1,3-benzoxadiazol-7- yl)aminoethyl]trimethylammonium; NBD-TMA (124)) and selected with 1 mg/ml geneticin (Gibco, Rockville, MD). Clones that continued to accumulate NBD-TMA were tested for transport of [14C]tetraethylammonium ([14C]TEA), and the clones that displayed the highest rate of TEA uptake were characterized in greater detail. The time

course and kinetics of TEA transport was measured in experiments using several

successive passages of each cell line. Values for Jmax and Kt for TEA transport were consistently similar for the individual clones of each transporter. Consequently, representative cells lines expressing either hOCT2 or rbOCT2, were selected to conduct the subsequent experiments in this study.

Transport experiments

CHOhOCT2 and CHOrbOCT2 cells were seeded in 12-well plates (USA Scientific,

Ocala, Fl) and grown to confluency. Once confluent (typically 24-48 hrs), transport

experiments were conducted by aspirating the media and preincubating each well of cells

in two successive 15 min exposures to 1 ml of Waymouth buffer (WB; in mM: 135 NaCl,

13 Hepes, 2.5 CaCl2-2H2O, 1.2 MgCl2, 0.8 MgSO4-7H2O, 5 KCl, 28 Glucose).

Following the preincubations, 400 µl of ‘transport buffer’ containing (typically) 0.4

µCi/ml [14C]TEA (~10 µM) and, in some cases, increasing concentrations of a test inhibitor in WB, were added to the wells. At intervals, the transport buffer was removed and each well was rinsed three times with 2 ml ice-cold WB containing 250 µM tetrapentylammonium to stop transport. Cells were solubilized with 400 µl 0.5 N NaOH in 1% SDS by shaking for 30 minutes. Solubilized cells were neutralized with 200 µl 1

N HCl, the solution triturated, and 500 µl removed and placed in a scintillation vial. The amount of radioactivity in each sample was determined using scintillation spectrometry

(Beckman model LS3801). Individual transport observations were performed in triplicate for each experiment, and observations were confirmed, typically 2 or 3 times

(i.e., n=2 or 3), in separate experiments using different cell passages. All experiments

were performed between passages 4 – 40 (post cloning), with no appreciable difference in

the results obtained with early vs. later passages (i.e., little change in Kt for TEA).

Modeling with Cerius2

This work was performed by Dr. Sean Ekins.

The computational molecular modeling studies were carried out as described in

more detail previously (31) using a Silicon Graphics Octane workstation (SGI, Mountain

View, CA). Molecular structures were used as either SMILES or sdf format and

imported into Cerius2 version 4.8 (Accelrys, San Diego, CA). Cerius2 QSAR was used to generate 54 descriptors including the default, Jurs descriptors, Shadow indices and

Octanol/water partition coefficients (ALogP98 and ClogP) for the molecules of the training and test sets. The forward stepwise regression method incorporated within

2 Cerius was then used to relate the log IC50 to a selection of these descriptors, and hence

result in a QSAR model. The model was validated for numerical stability and internal

consistency using both the Leave-One-Out (LOO) cross validation method and by

permuting, or randomizing, the response variable a number of times.

CoMFA

OCT2 substrates were assigned partial atomic (point) charges at neutral pH (7.4)

by performing a 1SCF MOPAC calculation using the AM1 Hamiltonian (keywords: EF,

PRECISE, MMOK). Molecules containing positive ionizable groups with a pKa>8.4 (i.e.,

>90% ionized at pH 7.4) were modeled in the charged state using the additional keyword

CHARGE=1. Molecules were superimposed using the FieldFit routine in Sybyl and imported into a molecular spreadsheet. CoMFA descriptors were used as independent variables, whereas the dependent variable (biologic descriptor) used in these studies was logIC50.

HIPHOP pharmacophore development

This work was performed in collaboration with Dr. Sean Ekins.

Pharmacophore models were constructed using Catalyst™ version 4.9 (Accelrys,

San Diego, CA) to generate a common features (HIPHOP) (43) pharmacophore for the selective inhibitors of rbOCT2 (cimetidine, guanidine, NBD-TMA, N1- hydroxyethylpyridinium) and hOCT2 (carbachol, tyramine, choline, nicotine, metformin and serotonin). TPrA, clonidine and TBA were comparatively selective inhibitors for rbOCT2 but were severely limited in the number of molecular features that could be used for successful pharmacophore generation and were, therefore, excluded from our analysis.

The principal molecule for hOCT2 was carbachol, to which the other molecules were aligned, whereas for rbOCT2 the molecules were aligned to cimetidine. Substrate molecules were then aligned using hydrophobic, H-bond acceptor, H-bond donor and positive charge and positive ionizable features in the HIPHOP algorithm within

Catalyst™.

4.3 Results

Kinetics of TEA transport mediated by the human and rabbit orthologs of OCT2

Figure 4.1 shows the time course of [14C]TEA transport into CHO cells that stably

expressed either hOCT2 (Figure 4.1A) or rbOCT2 (Figure 4.1B). In both cases,

accumulation of labeled substrate increased with time in a near linear fashion for 5 min,

and was blocked ~95% by coexposure to 2.5 mM unlabeled TEA. Extrapolation of these

time courses to time zero resulted in positive intercepts. This did not represent non-

specific binding of labeled substrate to the cells or residual label left after rinsing, both of

which would have been revealed in the level of activity measured in the presence of 2.5

mM unlabeled TEA. These positive intercepts were only noticed in cells that expressed

transporter; accumulation of [14C]TEA into wild type CHO cells was the same low level as that noted in transporter-expressing cells when blocked by unlabeled TEA (data not shown). Busch et al (126) noted that uptake of MPP into HEK-293 cells that stably expressed hOCT2 occurs very rapidly, reaching steady state within 5-10 sec. However, in the present case, the ‘rapid’ uptake of TEA into OCT2-expressing CHO cells did not represent an approach to steady state, but was followed for many minutes by a continuous, time-dependent component of mediated transport. Whereas the mechanistic basis of the rapid component of [14C]TEA accumulation is not known, it had kinetic properties effectively identical to the fraction of total uptake that clearly represented time dependent cellular transport, as discussed below. Consequently, for the subsequent kinetic analyses, we used 5 min or 2 min uptakes of [14C]TEA to provide estimates of the initial rate of

TEA uptake into CHOhOCT2 or CHOrbOCT2, respectively.

Figure 4.2 shows the kinetics of TEA transport into CHOhOCT2 (Figure 4.2A) or

CHOrbOCT2 (Figure 4.2B). For both processes, the addition of unlabeled TEA inhibited uptake of [14C]TEA by a process adequately described by the Michaelis-Menten equation

for competitive interaction of the labeled and unlabeled substrate (127):

* J max [ T] J = * + C eq. 2 K t +[ T] + [T]

where J is the rate of [14C]TEA transport from a concentration of labeled substrate equal to [*T]; Jmax is the maximum rate of mediated TEA transport; Kt is the TEA concentration that resulted in half-maximal transport (Michaelis constant); [T] is the concentration of unlabeled TEA in the transport reaction; and C is a constant representing the component of total TEA uptake that was not saturated (over the range of substrate concentrations tested) and presumably reflected the combined influence of diffusive flux, non-specific binding and/or incomplete rinsing of the cell layer. In 3 separate experiments, the Kt values for TEA transport mediated by hOCT2 or rbOCT2 were 47.2

± 2.2 µM and 80.9 ± 13.3 µM, respectively (with Jmax values of 8.0 ± 2.0 and 29.6 ± 18.9 pmol cm-2 min-1, respectively). As alluded to earlier, analysis of the residuals determined from fitting equation 1 (127) to the kinetic data suggested that the saturable component of

TEA transport was adequately described by the activity of a single, hyperbolic process.

The same observation was evident in the (upcoming) analyses of inhibitor interaction with these transporters. We interpreted this as indicating that the rapid, displaceable binding component of OC accumulation in these cells involved kinetics very similar to that of the time-dependent element of OC uptake and, consequently, would have little effect on the calculation of the kinetic constants Kt, Ki or IC50.

Inhibition of OCT2 activity by selected organic electrolytes

In an attempt to develop a model of the physical and structural basis of substrate-

transporter interaction for OCT2, we assessed the kinetics of inhibition of [14C]TEA

transport produced by an array of potential substrates for human and rabbit OCT2. With

the exception of several anionic inhibitors, the kinetics of inhibition were well described

by the relationship (128):

J [T*] J = app + C eq. 3 Kapp + [I]

where Japp is defined as (Ki/Kt)Jmax, [I] is the concentration of the test agent, and

Kapp is an apparent inhibitory constant (Ki) for the test agent that is defined as Ki(1 +

[T*]/Kt). When [T*] is << Kt, Kapp ≈ Ki. The application of this equation carries the tacit

assumption that the inhibitory interactions observed are competitive in nature and reflect

binding of substrate and inhibitor at a common binding site. Although that is

demonstrably the case for certain compounds (e.g., cimetidine, tyramine, NBD-TMA;

(117, 124) and unpublished observations), and reasonably assumed to be the case for

others (owing to marked structural similarities with molecules known to be OCT2

substrates), we have not rigorously proven this to be the case for all compounds used in

this study. Consequently, we will henceforth refer to the kinetic constants calculated

through application of equation 2 as ‘IC50’ values.

The test agents used here were selected to represent a broad range of the parameters suspected of influencing binding to the transport site of OCT2, including hydrophobicity (e.g., ClogP and ALogP98), molecular weight, basicity, and 3D

configuration. In addition, we considered it important to have compounds for which

OCT2 displayed a broad range of apparent affinities in order to generate QSAR models.

Figure 4.3 shows the effect of increasing the test agent concentration on the inhibition of

TEA transport mediated by either hOCT2 (Figure 4.3A) or rbOCT2 (Figure 4.3B) for

four representative compounds (ethylacridinium, clonidine, tyramine, and guanidine)

with IC50 values that spanned 5 orders of magnitude. Table 4.1 lists the IC50 values for the compounds included in the training and test sets examined in this study. Interestingly,

these orthologous transporters displayed both remarkable similarities in their apparent

affinities for selected compounds, and marked differences. For example, whereas human

and rabbit OCT2 had virtually identical IC50 values for ephedrine (Figure 4.4A), hOCT2

had a 10-fold higher apparent affinity for carbachol (than rbOCT2; Figure 4.4B), while

rbOCT2 had a 20-fold higher apparent affinity for cimetidine (than hOCT2; Figure 4.4C).

Figure 4.5 compares the IC50 values measured for the battery of test agents against TEA

transport mediated by hOCT2 (x-axis) and rbOCT2 (y-axis). Although it is evident that

there was a marked correlation between increasing IC50 values measured for inhibition of hOCT2-mediated TEA transport and parallel increases measured for rbOCT2, on average,

IC50 values measured for the rabbit ortholog were approximately 50% lower than those

measured for the human ortholog. Nevertheless, as suggested by the data presented in

Figure 4.4, there were a number of exceptions to this general rule. Figure 4.6 shows the

ratio of IC50 values measured for hOCT2- vs. rbOCT2-mediated transport. Whereas the

rabbit OCT2 ortholog displayed a significantly greater affinity (than the human) for 9 of

28 compounds tested, eight of 28 compounds showed the opposite, i.e., hOCT2 displayed

a greater affinity for them than did rbOCT2.

As previously mentioned, a positive correlation between hydrophobicity and

affinity has been reported for the interaction of substrates and inhibitors of renal OC

transporters (e.g., (121)). Figure 4.7 shows the relationship between IC50 and ALogP98 of the diverse group of inhibitors of hOCT2 employed in the present study. Although there was a significant correlation between these parameters, it was comparatively modest (r2 = 0.38) and numerous ‘outliers’ were evident. It is also relevant to note that a plot of IC50 versus another commonly used calculated hydrophobicity indicator, ClogP,

suggested that there was no significant correlation between these parameters (data not

shown). This somewhat unexpected observation reflects the fact that the commonly used

algorithms for calculation of octanol:water partition coefficients frequently show rather

modest agreement with one another, as shown by the comparatively weak correlation (r2

= 0.27) between AlogP98 and ClogP values for the compounds used in the present study

(data not shown). We raise this issue because it underscores the desirability of developing a more precise means to predict the relationship between substrate structure and binding to OCT2 rather than using a predicted measure of hydrophobicity alone.

Generation of QSAR Models for hOCT2 – 2D-QSAR (Cerius2)

The comparatively weak correlation between substrate/inhibitor hydrophobicity

and the measured interaction with hOCT2 led us to consider a more rigorous method for

developing a predictive model of substrate-transporter binding. We have previously used

Cerius2 to develop a descriptor-based QSAR model of substrate binding to hOCT1 (31),

an approach that proved to be superior to one based on the use of Catalyst to develop a

3D pharmacophore of binding to the transporter. Consequently, a descriptor-based 2D-

QSAR model for hOCT2 was built using a small selection of the molecular descriptors

generated by Cerius2. The following equation, produced using forward stepwise regression, incorporates the five molecular descriptors of the training set of 30 molecules that proved to be most strongly correlated with inhibition of hOCT2 activity:

Log IC50 = -0.925378 + 0.125798 * Rotatable bonds - 0.412128 * ALogP98 +

4.05786 * Jurs-RNCG + 1.62335 * Jurs RPSA + 0.02947 * Shadow YZ eq. 4

In this equation RNCG represents the charge of the most negative compound

divided by the total negative charge, RPSA is the relative polar surface area, and shadow

YZ is the area of projection in the YZ plane. Figure 4.8A shows the correlation between

hOCT2 the hOCT2 IC50 values measured using the CHO cells and those predicted by the

Cerius2 model; the model yielded an r2 = 0.92, leave one out q2 = 0.74, and F-test = 20.9 for the compounds comprising the training set. The model was randomized 9 times to give 90% confidence (r2 = 0.49 ± 0.09) that represents 4.25 standard deviations from the original model described in equation 4. Interestingly, Cerius2 was, however, unable to converge on a model describing binding of the training set molecules to rbOCT2 using these same descriptors.

The model outlined in equation 4 was used to predict for hOCT2 the IC50 values

for a test set of six diverse compounds selected to reflect the structural diversity

associated with the training set. Although the predicted IC50 of one of the six test

compounds (ibuprofen) was poorly predicted (the open circles in Figure 4.8A), predicted-

versus-measured IC50s for the remaining five compounds resulted in an r2 = 0.68.

Inhibition studies with N-1 substituted pyridiniums and quinoliniums with hOCT2

A set of N-1 substituted pyridiniums containing a phenyl substituent at the 3 or 4

position was investigated to determine if the 3D placement of this hydrophobic mass

influenced binding to hOCT2. The N-1 substituent was also varied to increase the

compounds hydrophobicity (hydroxyl ethyl< ethyl< benzyl). Table 4.2 shows the ALogP

and IC50 values generated for each compound. The data show that as the hydrophobic phenyl ring was rotated about the pyridinium, there was no change in affinity for the transporter. However, as hydrophobicity of the N-1 substituent was increased, the affinity for the transporter was also increased. In addition, a set of quinoliniums, also containing the differing N-1 substituents mentioned above, were tested. Again, affinity was positively correlated with hydrophobicity, and IC50 values were very similar to those for the corresponding 3 or 4 phenylpyridinium. The correlation between log IC50 and

AlogP and CLogP for these 9 compounds was r2= 0.85 and 0.57, respectively, while the

Cerius2 2D-QSAR predicted these same molecules with r2 = 0.70. In a previous study

(31) these two sets of compounds were used to perform inhibition kinetics of TEA in

HeLa cells stably transfected with hOCT1; IC50 values generated in this study are listed

in Table 4.2. As with hOCT2, a decrease in the IC50 values for interaction with hOCT1

was correlated with an increase in hydrophobicity for all subsets of compounds.

However, unlike the situation observed for hOCT2, as the hydrophobic mass was rotated

around the pyridinium, IC50 for interaction with hOCT1 increased, suggesting spatial

arrangement of hydrophobic mass effects a compound’s interaction with hOCT1 more

substantially than with hOCT2.

When all the test set molecules were combined with the training set a total of 45

molecules were available for model building. Using stepwise regression the r2 decreased

to 0.70, q2 = 0.55, F test = 18.4.

Log IC50 = 3.07539 + 0.202466 * Rotatable bonds -0.306335 * AlogP98 –

0.00051953 * JURS-DPSA-2 - 0.0037616 * Jurs-TASA + 0.168926 CLogP eq. 5

The model was randomized 9 times to give 90% confidence the r2 = 0.42 ± 0.10

which represents 4.07 standard deviations from the original model equation 4. It would

appear that adding this data caused deterioration in the model statistics, probably due to

the inclusion of the phenylpyridinium compounds for which hydrophobicity appears to be

disproportionately important for binding. The importance of hydrophobicity is evident in

the inclusion of AlogP, CLogP and the Jurs descriptors (TASA, Total hydrophobic

surface area; DPSA-2, Difference in total charge weighted surface areas). Figure 4.8B

shows the relationship between the experimentally determined IC50 values for the 45 test compounds and the values predicted using equation 5. Exclusion of three anionic compounds (PAH, probenecid and ibuprofen; gray diamonds) increased the correlation between measured and predicted (r2 = 0.81).

3D-QSAR (CoMFA)

A FieldFit alignment based on manually selected overlapping points was used as a basis to generate CoMFA models for human and rabbit OCT2. Both transporters displayed the greatest apparent affinity for ethylacridinium (IC50 values of 90 and 40 nM

for hOCT2 and rbOCT2, respectively; Table 4.1), so it was selected as the template

molecule for overlapping all other OCs. The positively charged nitronium and the center

92 of hydrophobic mass were used as two features to guide the alignment. A force constant of 20 was used for FieldFit. All aligned structures were relaxed and then submitted to

CoMFA analysis. Test set compounds (clonidine, guanidine, tetramethylammonium, N-1- hydroxyethylpyridinium, and 2,4-dimethyl-9-H-indenol1[2,1-c]pyridinium) were randomly selected while maintaining the coverage of training set activity. The following fields were generated for each CoMFA model: CoMFA_standard; CoMFA_indicator;

CoMFA_parabolic; CoMFA_Hbond; CLogP; and Molconn-Z. The correlation between each of these descriptor fields and measured IC50 values was calculated and compared.

Pindolol and ibuprofen were identified as outliers both by Factor Analysis and CoMFA runs and, so, were excluded from both CoMFA models. The QSAR statistics for the best correlation are listed in Table 4.3.

CoMFA contours at 80% confidence levels were generated for each model and these are shown in Figure 4.9. A large blue contour covering the positive center (mainly ammonium) suggests an important role for positive charge at this position. The small green contour over the phenol ring indicates the necessity of a sterically bulky group at this position. These two features should be expected because of the way our alignment was set up. Quite interestingly, a red contour next to the green contour appeared in both human and rabbit CoMFA models. This suggests a correlation between electronegative charge in the area and higher affinity to OCT2. Furthermore, this could indicate that a small negative charge or delocalized point charge close to the positive center might play a stabilizing role in the binding of substrates to OCT2. Figure 4.10 shows the relationship between predicted vs. measured log IC50 values for the training and test sets for human and rabbit OCT2. The predictive value of the CoMFA models is evident in the r2 values

93 of 0.97 for both training sets, and r2 values of 0.85 to 0.89 for the test set for human and rabbit OCT2, respectively. It should be noted that both models had difficulty predicting

2,4-dimethyl-9-H-indenol[2,1-c]pyridinium (residual values 3.73 and 5.63 µM for hOCT2 and rbOCT2, respectively), most likely caused by a slight misaligned overlap of its nitrogen group out of plane from the position where the other molecules’ electropositive moieties reside (data not shown).

OCT2 HIPHOP pharmacophore development

The common feature alignment of four selective rbOCT2 inhibitors suggested a pharmacophore that was characterized by a positive charge feature and a H-bond donor at a distance of 5.89 Å with an angle of 129.97o between the positive charge and the H-bond donor vector (Fig 11A shows alignment of cimetidine within the pharmacophore).

Ephedrine was used as an example of a molecule with no disproportionate selectivity for either transporter; it shows intermediate mapping to both pharmacophores (Fig 11B and

11C). The alignment of the six most selective hOCT2 inhibitors suggested a pharmacophore with a positive charge feature and a H-bond acceptor feature at a distance of 5.72 Å with an angle of 92.76o between the positive charge and the H-bond acceptor vector (Fig 11D shows alignment of carbachol within the pharmacophore).

4.4 Discussion

In light of the increasing pharmacological significance of renal secretion as a defining factor in the bioavailability of a vast array of cationic drugs (129, 130), there is obvious value to the ability to predict the extent to which potential substrates interact

with key elements in the renal secretory pathway. Ullrich made a detailed study of the

structural specificity of basolateral organic cation transport in rat kidney using the in vivo stopped-flow capillary microperfusion technique (121, 131). Interpretation of their observations is complicated by the current knowledge that, in rat kidney, basolateral OC transport involves at least two transporters, i.e., OCT1 and OCT2, which have, for at least some substrates, very different selectivity characteristics (132, 133). Consequently, the general rules concerning the physicochemical factors that influence substrate binding to

OCTs must be viewed as an average response of the interaction with multiple transporters operating in parallel. Nevertheless, the principal factor influencing substrate interaction with OCTs was found to be hydrophobicity, in concurrence with some of the earliest studies on renal OC secretion (134). In the present study there was a correlation, albeit weak (r2 of 0.38), between hydrophobicity and the interaction of our training set with hOCT2. A similar observation was noted in a recent examination of the factors that influence substrate binding to hOCT1 (31). Although within a group of structurally related compounds binding efficacy can be more closely correlated with hydrophobicity as a single criterion for interaction with OCTs (135), (136). It is evident, and not surprising, that factors other than the oil:water partition coefficient (or predictors of this property) play key roles in stabilizing substrate binding to OCTs.

The present study is, to the best of our knowledge, the first to apply a combination of in vitro and computational approaches to identify factors other than hydrophobicity that play a role in defining substrate interactions with OCT2. Our effort to develop a strategy for predicting molecular criteria that influence binding to OCT2 involved the use of two different computational methods, namely, Cerius2 to develop a 2D-QSAR model

of substrate interaction and CoMFA to develop a 3D-QSAR model. The latter, in

addition to predicting binding efficacy also held the promise of providing visual

information on the structural determinants that influence substrate-transporter interaction.

The Cerius2 models (equations 4 and 5) did underscore the importance of hydrophobicity as a determinant in influencing substrate binding to hOCT2, as they were dominated by parameters such as AlogP98, CLogP, and Jurs descriptors. However, a role for steric factors was also suggested in equation 4 by the inclusion of a Shadow parameter (Shadow

YZ; area of the molecular shadow in the YZ plane). Interestingly, our previous study on characteristics of binding hOCT1 identified Shadow nu (ratio of largest to smallest dimension) as a principal determinant of substrate interaction for this transporter (31). In the previous study, we noted that the placement of planar hydrophobic mass about a pyridinium nucleus exerted a systematic effect on binding to hOCT1 (i.e., 4- phenylpyridinium compounds interacted with substantially lower IC50 values than did, for example, quinolinium compounds), consistent with the conclusion that the hOCT1 binding sites favors binding of ‘longer’, rather that ‘wider,’ molecules. In contrast, in the present study, we found no such systematic effect with respect to the binding of 4- phenyl-, 3-phenyl, vs. quinolinium compounds (Table 4.2), and this was reflected in emphasis on binding of a bulk area term (Shadow YZ) rather than a term that emphasized the binding efficacy of long, narrow substrates. The presence of the rotatable bond descriptor suggests that hOCT2 may also prefer flexible substrates, a characteristic not implicated in hOCT1 binding. The degree of substrate ionization indicated as important for hOCT2 substrate binding was not evaluated in these computational studies (123). It should also be noted that both the 2D and 3D analyses employed here implicitly assume

that the binding site/region of OCT2 has similar characteristics when exposed to

extracellular and cytoplasmic aspects of the membrane and, for at least some substrates

(e.g., tetrabutylammonium and corticosterone) this has been shown not to be the case

(137). Consequently, the models describe a hypothetical binding site with mixed

properties of the extracellular and intracellular oriented conformations of the transporter.

The 3D-QSAR that resulted from application of CoMFA served to emphasize two

important issues concerning the influence of molecular size/shape on binding of substrate

to OCTs. First, it underscored the observation, suggested by the 2D-QSAR, that steric

factors clearly play a role in the binding process. This is evident in the marked

improvement in predictive power of the 3D-QSAR (r2 of 0.97; Figure 4.10) compared to

the 2D-QSAR (r2 of 0.8 to 0.9; Figure 4.8). The second issue evident from the CoMFA,

interestingly, was the multispecificity of the binding site evident from the substrate

overlays within the binding region (Figure 4.9). Although the CoMFA contour plot

indicates that binding is enhanced or reduced by, for example, the presence of steric mass

in particular positions (i.e., green vs. yellow contours, respectively), it is also evident that

the OCT2 binding site (both for human and rabbit) is extremely permissive with respect

to the presence and placement of, in particular, hydrophobic moieties that radiate away

from positively charged binding center. The capacity to interact effectively with such a

structurally diverse set of compounds has important implications for the structural nature

of the OCT2 binding site. This clearly represents a challenge for future efforts to model

the binding site by identifying structural features that permit comparatively high affinity

interactions with molecules as structurally diverse as, for example, MPP (IC50 of 2.4 µM) and crystal violet (IC50 of 2.6 µM) (Table 4.1).

Comparison of the selectivity of the human and rabbit orthologs of OCT2 is also instructive. Whereas the two orthologs displayed very similar apparent affinities for many of the test agents, it was also evident that for just as many compounds their selectivities differed substantially (Figure 4.5). Importantly, these differences were not systematic: for some compounds, e.g., carbachol and tyramine, hOCT2 displayed a 5-15- fold higher affinity, whereas for other compounds, e.g., cimetidine and clonidine, it was the rabbit ortholog that displayed a higher (5-10-fold) affinity (Figure 4.6). The primary sequences of human and rabbit OCT2 are 96% similar (83% identical). The differences in selectivity between these two transporters presumably reflect structural differences, quite probably minor ones, the consequences of which include clearly distinct selectivity profiles. While this is not a surprising observation, it serves to emphasize that minor changes in a few amino acid residues can result in very significant changes in selectivity for some substrates, while having no effect on interaction with other substrates. For example, Gorboulev et al. (138) found that substituting a glutamate residue for the aspartate found at position 475 in rat OCT1 resulted in an 8-fold increase in apparent affinity for TEA, but had no effect on the interaction of the transporter with MPP.

Similarly, Leabman et al. (139) noted that single nucleotide polymorphisms (SNPs) in hOCT2 exerted substrate-specific effects on transport function. For example, the K432Q

SNP increased apparent affinity of the transporter for tetrabutylammonium and MPP, while having little or no effect on interaction of the transporter for metformin and quinidine. On the other hand, the A270S SNP decreased apparent affinity for MPP and tetrabutylammonium, while having no significant effect on interaction with metformin and quinidine. The differences in selectivity between the human and rabbit OCT2 also

emphasize the care that must be used when extending to humans observations obtained in

studies employing non-human orthologs.

Knowing that there are molecules that discriminate between the human and rabbit

orthologs of OCT2 provided an opportunity to probe the qualitative differences between

molecules with high affinity to the hOCT2 and rOCT2 transporters using a Catalyst

HIPHOP alignment. An alignment of the selective inhibitors for both transporters

indicated subtle but distinctive differences for recognition, which manifested in

differences in features and angles recognized for each transporter (Figure 4.11A-D). The

assessment of ephedrine, which shows no selectivity for either transporter, indicates that

this molecule can adequately map to both pharmacophores (Figure 4.11B,C). This

pharmacophore analysis provides valuable information that can be used for testing

subsequent homology models for both transporters by comparing the pharmacophores

and aligned selective inhibitors docked into the transporters. Even though the features on

these pharmacophores are similar, the approach is sensitive enough to identify a

difference in the orientation of the H-bonding features (> 37o). This could infer variability in the disposition of critical amino acids for interaction with inhibitors within the respective transporters.

In summary, computationally derived QSAR models of the basis of OCT2 selectivity were determined for the human and rabbit orthologs of the renal organic cation transporter, OCT2. A 2D-QSAR emphasized the importance of hydrophobicity as an important determinant in the binding of substrates to OCT2 alongside structural bulk and molecular flexibility. A 3D-QSAR displayed better predictive power and served to emphasize the fact that molecular size and shape plays a significant role in defining the

99 interaction of substrates. The CoMFA models also highlighted the multispecificity of the

OCT2 binding site by noting the importance of structural features in selected regions, and the permissiveness of the binding site with respect to steric bulk in other regions. Marked differences in the selectivity of the human and rabbit orthologs of OCT2 also underscored the fact that very modest differences in the amino acid residue composition of the protein can result in substantial changes in affinity of the respective transporters for some substrates while having no effect on the interaction with other substrates. The general hydrophobic substrate promiscuity of the OCT2 transporter draws immediate parallels with P-gp and other proteins of importance to drug discovery that can bind a diverse array of xenobiotics to ultimately aid in their elimination from the body (140).

100

Figure 4.1. Time course of TEA uptake into CHO cells stably transfected with either hOCT2 or rbOCT2. Each point is the mean (±SE) of triplicate measures of 16 µM [14C]TEA uptake at each time point, determined in a single representative experiment. Uptake was measured in the absence and presence of 2.5 mM unlabeled TEA.

101

Figure 4.2. Kinetics of TEA transport in CHO cells stably transfected either hOCT2 or rbOCT2 .Five minute (hOCT2) or 2 min (rbOCT2) uptakes of 7.2 µM [14C]TEA were measured in the presence of increasing concentrations of unlabeled TEA (0-2500 µM). Each data point is the mean (± SE) triplicate measures of uptake in a single experiment. For the representative experiments shown, the Kt and Jmax for TEA uptake was, respectively 48 µM and 11.6 pmol cm-2 min-1 for hOCT2, and 67 µM and 14.4 pmol cm-2 min-1 for rbOCT2. Lines were fit to the data using a non-linear regression algorithm (SigmaPlot 3.0).

102

Figure 4.3. Effect of increasing concentration of several test inhibitors on uptake of [14C]TEA mediated by either hOCT2 or rbOCT2. Each point is the mean (±SE) of uptakes measured in two or three separate experiments with each test compound. IC50 values were (for hOCT2 and rbOCT2, respectively): ethylacridinium [z], 0.09 and 0,04 µM; clonidine [], 2.2 and 0.19 µM; tyramine [S], 106 and 426 µM; and guanidine [], 2300 and 272 µM.

103

Figure 4.4. Relative effect on increasing concentrations of ephedrine (A), carbachol (B) or cimetidine (C) on uptake of [14C]TEA mediated by either hOCT2 [z] or rbOCT2 [{]. Each point is mean (±SE) of uptakes measured under each condition measured in two separate experiments. Data were normalized to the level of uptake measured in the absence of the test inhibitor. Lines were fit to the data using a non-linear regression algorithm (SigmaPlot 3.0). IC50 values are listed in Table 4.1.

104

14 Figure 4.5. Comparison of IC50 values for inhibition of [ C]TEA transport mediated by hOCT2 (x axis) or rbOCT2 (y axis). Each point is the mean of two or three IC50 values measured in cells expressing either the human or rabbit ortholog of OCT2 (±SEs). The solid line is the linear regression of the logs of the measured and calculated parameters; the dotted line is the line of unity.

105

Figure 4.6. Comparison of the relative effect of each test compound as an inhibitor of hOCT2 versus rbOCT2. The length of each bar represents the ratio of a test agent’s IC50 for inhibition of TEA transport mediated by hOCT2, to that for transport mediated by rbOCT2. Error bars reflect the sum of the SEs determined for each agent as an inhibitor of the human and rabbit transporters. Asterisks designate those ratios that differed from unity by more than 2x the sum of the individual SEs.

106

Figure 4.7. Relationship between the IC50 for inhibition of TEA transport mediated by hOCT2 and that compound’s calculated oil:water partition coefficient (expressed as ALogP98, determined by Cerius2). The solid line is the regression of these two parameters; the dotted lines indicate 95% confidence limits.

107

2 Figure 4.8. Correlation between Cerius predictions and actual LogIC50 (A) Relationship 2 for hOCT2 between the IC50 values predicted using the 2D-QSAR (Cerius ) outlined in

eq. 4 (see text) and the measured IC50 values for the training set of 30 compounds (●) and the test set of six compounds (○). The solid line represents the regression of predicted vs. measured IC50 for the training set (r2 = 0.92). The major outliers of the test set, (from left to right) were 3,5-dimethyl-4-phenylpyridine-1-oxide, ibuprofen, famotidine, and amantidine. (B) Relationship between the IC50 values using the 2D- 2 QSAR (Cerius ) outlined in eq. 5 (see text) and the measured IC50 values for the complete set of cationic compounds employed in this study (45 molecules - ●). The three anionic compounds studied (PAH, probenecid and ibuprofen; ♦) were excluded from the analysis as they represent outliers. The solid line represents the regression of predicted vs. measured IC50 for the training set (r2 = 0.81; dotted lines represent 95% confidence limits).

108

Figure 4.9. Structural overlap of the OCT2 compounds and CoMFA contour maps for (A) hOCT2 (q2, 0.60) and (B) rbOCT2 (q2, 0.53). Red and blue contours are visualized at the 20% standard coefficient level, representing those areas surrounding the molecules where electronegative and positive charge, respectively, significantly contribute to OCT2 IC50 values. Analogously, green and yellow contours indicate the fields where steric bulk significantly increases or decreases OCT2 IC50 values, respectively.

109

Figure 4.10. Relationship between CoMFA predicted and measured LOGIC50 (hOCT2 (A) and rbOCT2 (B)) (training set: ●; test set: ○). The solid line represents the regression of predicted vs. measured IC50 for the training set (r2 = 0.97; dotted lines represent 95% confidence limits). For hOCT2 the leftmost test set outlier was 2,4- dimethyl-9-H-indenol[2,1-c]pyridinium; the right most outlier was guanidine. For rbOCT2, the leftmost outliers were clonidine and 2,4-dimethyl-9-H-indenol[2,1- c]pyridinium.

110

Figure 4.11. HIPHOP pharmacophores for rbOCT2 and hOCT2. A: Cimetidine mapped to rabbit HIPHOP model, B: Ephedrine mapped to rabbit HIPHOP model, C: Ephedrine mapped to human HIPHOP model, D: Carbachol mapped to human HIPHOP model

111

hOCT2 rbOCT2

Compound IC50 IC50 ( M) ( M) Training Set Ethylacridinium 0.09 ± 0.03 0.04 ± 0 2,4-DIPyr 0.34 ± 0.20 0.40 ± 0.03 1,3,5-TPPyr 0.77 ± 0.07 0.61 ± 0.06 Clonidine 2.2 ± 0.51 0.19 ± 0.02 MPP 2.4 ± 0.04 1.4 ± 0.17 Crystal Violet 2.6 ± 1.20 2.05 ± 0.15 Tetrapentylammonium 10.5 ± 3.09 10.4 ± 5.2 NBD-TMA 13.5 ± 2.50 2.3 ± 0 Phenformin 15.3 ± 1.86 7.3 ± 1.5 Tetrabutylammonium 19.5 ± 1.50 2.9 ± 0.26 Tetrapropylammonium 20.0 ± 0.58 1.6 ± 0.29 Nicotine 22.2 ± 3.05 55.2 ± 5.10 Ephedrine 29.0 ± 1.00 31.0 ± 3.00 Ranitidine 39.7 ± 16.10 27.0 ± 5.00 TEA 46.3 ± 1.89 86.5 ± 20.50 Pindolol 50.5 ± 3.50 67.5 ± 10.50 4-Phenylpyridine 57.5 ± 16.50 37.5 ± 3.50 Cimetidine 70 ± 10.00 3.3 ± 0.33 Tyramine 106 ± 2.0 426 ± 8.0 1-(2-Hydroxyethyl)pyridinium 112 ± 17.5 42.0 ± 3.00 Carbachol 248 ± 23.0 1439 ± 89 N1-Methylnicotinamide 303 ± 15.4 180 ± 13.0 Serotonin 310 ± 18.0 664 ± 44 Metformin 339 ± 5.3 808 ± 94 Choline 381 ± 75.6 1388 ± 401 Tetramethylammonium 525 ± 109 850 ± 44 Guanidine 2300 ± 536 272 ± 16 Histamine 3251 ± 497 427 ± 135

Test Set Cyclohexylamine 8.2 ± 0.80 18.0 ± 1.00 3,5-Dimethyl-4-phenylpyridine- 1-oxide 11.9 ± 3.15 47.0 ± 12.00 Amantadine 19.7 ± 3.56 31.7 ± 5.00 Procainamide 57.5 ± 1.50 30.0 ± 9.0 Famotidine 111 ± 38.0 7.4 ± 1.20 Ibuprofen ~10 mM 3100 ± 500

Table 4.1. Observed IC50 values for hOCT2 and rbOCT2 for structurally diverse organic cations. Each IC50 is the mean value of at least two separate experiments

112

Compound IC50 (µM) AlogP hOCT2 hOCT1 1-(2-Hydroxyethyl)-4-phenylpyridinium 10.3 ± 1.9 16.2 ± 1.4 1.76 1-(2-Hydroxyethyl)-3-phenylpyridinium 10 ± 2.3 31.1 ± 2.3 1.76 1-(2-Hydroxyethyl)quinolinium 14 ± 0.9 80.6 ± 3.2 1.58 1-(Phenyl)methyl-4-phenylpyridinium 2 ± 0.1 9.3 ± 0.6 3.88 1-(Phenyl)methyl-3-phenylpyridinium 0.8 ± 0.1 5.5 ± 0.9 3.88 1-(Phenyl)methylquinolinium 0.47 ± 0.06 14.3 ± 0.7 3.70 1-Ethyl-4-phenylpyridinium 4.9 ± 0.8 7.1 ± 1.1 2.65 1-Ethyl-3-phenylpyridinium 4.45 ± 0.7 28.3 ± 4.5 2.65 1-Ethylquinolinium 3.3 ± 0.5 67.6 ± 11.2 2.47

Table 4.2. IC50 values for the inhibition of TEA by the phenylpyridiniums and quinoliniums in CHO cells stably transfected with hOCT2 (this study) and HeLa cells stably transfected with hOCT1 (Bednarczyk et al., 2003). Each IC50 is the mean value of at least 2 experiments (described in Materials and Methods). ClogP values calculated with Daylight software using Cerius2 (Accelrys, San Diego, CA).

113

No. of PRESS Q square components r square s %Steric %Electrostatic hOCT2 0.71 0.596 4 0.973 0.182 0.65 0.35 rbOCT2 0.918 0.531 4 0.972 0.224 0.65 0.35

Table 4.3. CoMFA statistics for human and rabbit OCT2 models

114

CHAPTER 5

COMPARATIVE PHARMACOPHORE MODELING OF ORGANIC ANION

TRANSPORTING POLYPEPTIDES:

A META-ANALYSIS OF RAT OATP1A1 AND HUMAN OATP1B1

5.1 Introduction

The organic anion transporting polypeptides (OATP) are key membrane bound transporters expressed in many organs including intestine, liver, lung, choroid plexus, blood brain barrier and other organs (141). This transporter family mediates sodium- independent transport of a diverse array of molecules that are mostly anions as well as organic cations, steroid conjugates, organic anions and xenobiotics (142, 143) by coupling uptake with the efflux of bicarbonate (144) or glutathione (145). The OATPs share some substrate overlapping specificity with other promiscuous efflux transporters such as P-gp and MRP2, indicative of a degree of coordination. As a result, OATPs have been implicated in drug-drug interactions (146), as exemplified by the interactions between cerivastatin and cyclosporine A (147) as well as cerivastatin, gemfibrozil and its

115

glucuronide metabolite (148). Thirty six mammalian OATPs have been identified, but

only a few of these have been characterized in any detail. Currently, 11 human OATPs

have been identified (Figure 5.1) and a new species independent nomenclature system

has been proposed for all of the OATPs (SLCO) (143). OATP1B1 (previous names

OATP-C, LST-1, OATP2, SLC21A6), consisting of 670 amino acids with 12 putative

membrane spanning domains, represents the most studied human OATP to date (149),

and is expressed on the basolateral plasma membrane of hepatocytes. This transporter has

been well characterized when transfected in Xenopus laevis oocyte and other expression

systems such as CHO, HeLa and Hek-293 cells. Many diverse molecules with a range of

Km values are substrates of this protein, including pravastatin, dehydroepiandrosterone sulfate (DHEAS) and bromosulfophthalein (BSP) (Table 5.1).

An equally well studied rat SLCO namely, Oatp1a1 (Oatp1, Oatp, Slc21a1) that is expressed in the liver, kidney and choroid plexus appears to share many of the same substrates as OATP1B1 (67% identity) as well as organic cations (Table 5.2). The considerable substrate overlap of these rat and human transporters, though by no means identical (150), could suggest some degree of structural homology in key substrate recognition areas of the proteins that may require assessment using site-directed mutagenesis or other methods. As no high-resolution structures are presently available for these transporters, it is necessary to gain structural insight from alternative methods. If sufficient binding data are generated in vitro, these can be used to build a three- dimensional quantitative structure activity relationship (3D-QSAR) models (108). A recent report has briefly discussed 3D-QSAR studies of Oatp/OATP substrates which apparently produced a pharmacophore containing 2 H-bond acceptors, one H-bond donor

116

and two hydrophobic regions, although no model was shown or further details provided

(143). It is likely that computational modeling techniques can provide structural insights

into the molecular features of substrates that are common to these OATP transporters;

alternatively, they may indicate those structural features that differentiate one OATP

from the other. In the present study, we generated pharmacophore models for rat Oatp1a1

and human OATP1B1 with published Km values from different experimental systems.

Meta-analysis of the individual models allowed us to provide an understanding of the key molecular features for substrate-OATP transporter interactions.

117

5.2 Methods

Literature search

A complete literature search was performed up to September 2004 to retrieve

substrates for each OATP evaluated in this study. Various databases were consulted for

this task including PubMed (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi), the TP- search transporter database (http://www.ilab.rise.waseda.ac.jp/tp-search/) and ISI Web of knowledge (http://isiwebofknowledge.com/). Details from the publications were noted such as the cell system used for transfection, the Km value, and the species. Considerable care was taken to assign molecules to the correct OATP/Oatp. Comparisons between substrates with Km data shared between Oatp1a1 and OATP1B1 were used to derive a correlation (Table 5.3). In cases where multiple values for the same substrate could be found for each cell line, the mean value was then computed.

OATP pharmacophore development with Catalyst

The structure sketch and conformation generation were performed using the same settings as the previous hPepT1 study.

Ten hypotheses were generated using these conformers for each of the molecules in the training sets and the Km values, after selection of the following features:

hydrophobic, H-bond acceptor, H-bond donor, and the positive and negative ionizable

features for the substrates in the algorithm within Catalyst™. After assessing all 10

hypotheses generated, the lowest energy cost hypothesis was considered the best as this

possessed features representative of all the hypotheses and had the lowest total cost.

118

Validation of the Catalyst Models

The test sets contained molecules not included in the initial training sets as described previously. These test set molecules were fit by the fast-fit algorithm to the respective Catalyst™ models in order to predict a value as previously described for CYPs

(151). Fast fit refers to the method of finding the optimum fit of the substrate to the hypothesis among all the conformers of the molecule without performing an energy minimization on the conformers of the molecule.

Merging Catalyst Models

Pharmacophore models generated for different cell types for each transporter were merged in Catalyst using a pairwise comparison. Final pharmacophore models with the inter-feature distances labeled for clarity were illustrated using MOLEKEL 4.3, (Swiss

Center for Scientific Computing, Manno, Switzerland) (152).

Statistical evaluation of test set predictions

Observed and predicted data were graphed and fitted using KaleidaGraph

(Synergy Software, Reading, PA) to generate an r (correlation) value. The non parametric

Spearman Rho values were calculated as a measure of rank ordering in the test data using

JMP version 5.1 (SAS institute Inc, Cary, NC), where a value of 1 is optimal.

119

5.3 Results

Correlation between substrate data for rat Oatp1a1 and human OATP1B1

Overlapping substrate affinity for OATPs has been previously suggested due to

the relatively close phylogenetic proximity of members of the OATP family. We have

therefore made several comparisons using literature data published by many groups.

Comparisons between substrates with Km data shared between rat Oatp1a1 and human

OATP1B1 (Table 5.3) were used to derive a correlation (logKm, oatp1a1=0.55logKm, OATP1B1

+ 0.68). The analysis of all published data available to date across both of these transporters suggests 8 molecules in common which provides a reasonable correlation for

2 the Km values (r of 0.64).

Pharmacophores for human OATP1B1

Using the substrates described in the literature (Table 5.1), sufficient data were available from Hek-293 cells and Xenopus laevis oocytes expressing OATP1B1 to generate two separate pharmacophores. The OATP1B1-oocyte data set consisted of 12 molecules (Km range 0.0076 – 17 µM) in the training set. The pharmacophore generated in Catalyst contained 2 H-bond acceptors and three hydrophobes (Figure 5.1A) with an observed versus predicted correlation (r = 0.97) but minimal difference between the null and total cost values (Table 5.4). Similarly, after scrambling the molecules and Km data,

no significant differences were found in the statistics. A second pharmacophore,

OATP1B1-Hek consisted of 12 molecules (Km range 0.1 – 268 µM) in the training set.

The pharmacophore generated in Catalyst contained 2 H-bond acceptors and one

120

hydrophobe (Figure 5.1B) with an observed versus predicted correlation (r = 0.91) but

minimal difference between the null and total cost values. However after scrambling the

molecules and Km data the correlation decreased. Both pharmacophores compared favorably when merged, with good overlap of the H-bond acceptor features and central hydrophobic features (Figure 5.1F).

Pharmacophores for rat Oatp1a1

Among the data available on substrates described in the literature on oocytes,

CHO and HeLa cells expressing Oatp1a1 (Table 5.2), it was apparent that 3 separate pharmacophores could be generated. The Oatp1a1-oocyte data set consisted of 14 molecules (Km range 0.015 – 3300 µM) in the training set (s-dinitrophenyl-glutathione was excluded as it was an outlier). The pharmacophore generated in Catalyst contained 2

H-bond acceptors and three hydrophobes (Figure 5.1C), with an observed versus predicted correlation (r = 0.92) and a thirty points difference between the null and total cost values (Table 5.4). Similarly, a decrease in correlation and increase in the total cost were found after scrambling the molecules and Km data. A second pharmacophore,

Oatp1a1-CHO consisted of 12 molecules (Km range 3 – 3000 µM) in the training set. The pharmacophore generated in Catalyst contained 1 H-bond acceptor and two hydrophobes

(Figure 5.1E), with an observed versus predicted correlation (r = 0.90) but minimal difference between the null and total cost values. However, the correlation was found to decrease with no change in the total cost upon scrambling the molecules and Km data. A third pharmacophore, Oatp1a1-HeLa, consisted of 9 molecules (Km range 3.1 – 214 µM) in the training set. The pharmacophore generated in Catalyst contained 1 H-bond

121

acceptor, one hydrophobe and one negative ionizable feature (Figure 5.1F) with an

observed versus predicted correlation (r = 0.95) and larger total than null cost value.

However, no change was found after scrambling the molecules and Km data. The three pharmacophores compared favorably when merged, with good overlap of the H-bond acceptor features and central hydrophobic feature, while the negative ionizable feature was only present in one of three models which was generated with a narrower Km range

(Figure 5.1G).

Meta pharmacophores for rat Oatp1a1 and human OATP1B1

Due to the limited size of the training sets generated for each model with individual cell types, we attempted to increase the training set scope by combining data from multiple cell types. Combining data for the rat Oatp1a1 resulted in a training-set of

26 molecules (Km range 0.015 – 3300 µM). The meta-pharmacophore generated in

Catalyst contained 2 H-bond acceptors and three hydrophobes (Figure 5.1H) with a good observed versus predicted correlation (r = 0.90) and over a 30 point difference between the null and total cost values (Table 5.4). After scrambling the molecules and Km data the correlation decreased and the total cost increased to close to the null value. Combining data for the human OATP1B1 resulted in a training set of 18 molecules (Km range 0.0076

– 268 µM). The meta-pharmacophore generated in Catalyst contained 2 H-bond

acceptors and three hydrophobes (Figure 5.1I ) with an observed versus predicted

correlation (r = 0.92) and nearly a 20-point difference between the null and total cost values (Table 5.4). After scrambling the molecules and Km data the correlation decreased

and the total cost increased.

122

Conformational features of the meta-pharmacophores

The molecules BSP, cholate, DHEAS, estrone-3-sulfate and taurocholate exist in the training sets for meta-OATP1B1and meta-Oatp1a1. To illustrate that these molecules would map to the meta-pharmacophores, their relative positioning is shown in Figure 5.3.

DHEAS shows medium affinity to Oatp1a1, and low affinity to OATP1B1 This is validated in our OATP1B1 meta-pharmacophore models because the molecule fails to occupy a hydrophobic feature and H-bond acceptor (Figure 5.3A, C). DHEAS fits reasonably well to all Oatp1a1 pharmacophore features as expected from its relatively high affinity to Oatp1a1 (Figure 5.3B). The relative molecular dimensions of the pharmacophoric feature points is illustrated in Fig 3D-F. It should be noted that the

OATP1B1 pharmacophore that includes bilirubin occupies twice the dimensions of the oatp1a1 and OATP1B1 (without bilirubin) models.

External test sets for human OATP1B1 and rat Oatp1a1 pharmacophores

Additional molecules not included in the individual training sets for each cell type were used as test sets for the other respective model(s) for each transporter. The test set correlations using each single cell type OATP pharmacophore were as follows:

OATP1B1-Hek (r = 0.214, n = 7; Spearman rho -0.14, p = 0.76), OATP1B1-oocyte (r =

0.658, n = 7; Spearman rho 0.57, p = 0.18), Oatp1a1-oocyte (r = 0.412, n = 13; Spearman rho 0.39, p = 0.18), Oatp1a1-HeLa (r = 0.224, n = 18; Spearman rho -0.26, p = 0.28) and

Oatp1a1-CHO (r = 0.912, n = 15; Spearman rho 0.92, p = < 0.0001, Figure 5.2).

123

External Test Set Validation

It has been previously shown (153) that BSP-GSH is transported by rat Oatp1a1,

while enalaprilat, hippuric acid, benzoic acid, and harmol sulfate were not. These

molecules form an excellent external test for our Oatp1a1 pharmacophore model.

Hippuric acid and benzoic acid failed to fit to the model (indicative of them not being

substrates) and enalaprilat, and harmol sulfate resulted in modest to high Km values (69

µM and 550 µM respectively). BSP-GSH, the glutathione adduct of BSP, was predicted

to have a Km value of 23 µM, which indicates it is a good substrate, further validating the

Oatp1a1-meta pharmacophore model. Recently, troglitazone, its metabolites and similar compounds were shown to produce statistically significant inhibition of estrone-3-sulfate uptake in OATP1B1 (154). Indeed, when these molecules are fitted to the OATP1B1 meta model (without bilirubin) they are all predicted to have affinity for this transporter: troglitazone (predicted Km 7.9 µM), troglitazone glucuronide-M2 (1.6 µM), troglitazone sulfate-M1 (5.6 µM), troglitazone quinone-M3 (6.7 µM), pioglitazone (14 µM), and rosiglitazone (9.9 µM). The OATP1B1 inhibitor indocyanine green (155) was also fitted to this same pharmacophore, showing an overlap to the two hydrophobes and one H-bond acceptor (predicted Km 6.3 µM).

124

5.4 Discussion

As shown in the previous chapters as well as the following chapters,

computational approaches have been extremely useful in gaining insight into the ligand-

protein interaction in the absence of crystal structures of the transporters and receptors.

However, the quality and consistency of datasets have been a determining factor in the

overall predictive value of the QSAR models to date. It has been particularly challenging

to assimilate and model data acquired across species and experimental cell systems. As a

result, most QSAR studies have focused on datasets gathered from one species, cell type,

and frequently, one laboratory setting. In the present study we aimed to overcome these

issues by a combination of pharmacophore building and meta-analysis.

To our knowledge the application of QSAR models for OATP or Oatp has been

limited largely due to the absence of consistent datasets. A notable exception is a recent

study by Yarim and co-workers on rat oatp1a5 (156) who used CoMFA on 18 substrates.

They have identified similar chemical features that are responsible for OATP interaction;

one H-bond acceptor feature, one hydrophobic feature and a positive charge feature. An

improved understanding of the structural requirements of the OATPs may explain the

mechanisms underlying the reported drug-drug interactions as due to transporter

inhibition (146). As some studies have described multiple inhibitors of uptake with EC50 values (157), it may be possible to generate similar pharmacophores for inhibitors of the respective transporter. However, the difficulty in interpretation of whether these molecules are interacting with the same site or even sites responsible for transport is a disadvantage compared to modeling substrate Km data. It is therefore important to

125

determine the critical features of OATP substrates, in particular the extensively studied

rat Oatp1a1 and human OATP1B1.

Using the spectrum of substrates collated from the literature (Table 5.1), we built

2 pharmacophores with Hek-293 cells and X. laevis oocytes expressing OATP1B1

(Figure 5.1A,B). Both pharmacophores had high correlations and contained multiple H-

bond acceptors and hydrophobic features which overlapped when merged, suggesting

similarity (Figure 5.1F). Small differences between the null and total cost values (Table

5.4) before and after scrambling the molecules and Km data indicated these models may be limited however. Ideally upon scrambling one would expect the training set correlation to diminish as the affinity data is randomly assigned with molecular structure and the total cost for the hypothesis will be similar to the null hypothesis cost. This simple test can give some idea of whether the model derived originally is meaningful or likely to be similar to one generated with random data. Literature data for the Oatp1a1 substrates (Table 5.2) yielded 3 pharmacophores (Figure 5.1C-E). Interestingly, all models were subtly different but shared H-bond acceptors and hydrophobes, resulting in good overlaps in the merged pharmacophore. One Oatp1a1 pharmacophore contained a negative ionizable feature that likely relates to the most active molecules in the training set. Most of these models showed decreases in the training correlation and increases in the total cost after scrambling, indicating these models are of some utility. It is important to note that relatively small literature training sets (i.e., < 20 compounds) can dramatically impact the model cost statistics and their behavior upon scrambling.

Test sets were generated from the molecules not initially included in the individual cell type pharmacophores, but results were variable due to the limited nature of

126

the test sets. By combining data from different cell types we increased the training set

scope for each transporter model both in terms of the number of molecules and the

activity range. These so-called meta-pharmacophores for rat Oatp1a1 and human

OATP1B1 both contained 2 H-bond acceptors and three hydrophobes (Figure 5.1G-I)

although with different 3D coordinates (Table 5.5). Both models had larger differences

between null and total cost values with similar observed versus predicted correlations

compared with the individual models generated with separate cell lines. Also, after

scrambling the molecules and Km data for these meta-pharmacophores, the correlations decreased and the total cost increased for both models, indicating that they are statistically significant. It seems therefore, that the degree of correlation between substrates for these two transporters observed earlier could, in part, be due to the possession of similar molecular features necessary for interaction with each transporter, while the exact 3D coordinates differ to some extent. The key pharmacophore features for both of these transporters appear to be two H-bond acceptors at either end of a large hydrophobic area. These features will correspond with H-bond donors and a large planar hydrophobic recognition site on the respective transporters.

Five molecules (BSP, cholate, DHEAS, estrone-3-sulfate and taurocholate) exist in both training sets for meta-OATP1B1and meta-Oatp1a1. All but DHEAS have similar relative affinity to both transporters and map to both pharmacophores (Figs 3G-N). The fact that DHEAS shows medium affinity to Oatp1a1, and low affinity to OATP1B1 makes it a good substrate to compare both pharmacophore models. Hence, DHEAS was fitted to both pharmacophore models (Figure 5.3). A hydrophobic feature and H-bond acceptor was missed and explained the low affinity of DHEAS to OATP1B1 (Figure

127

5.3A). DHEAS fits reasonably well to all Oatp1a1 pharmacophore features and correlates

with its relatively high affinity to Oatp1a1 (Figure 5.3B). Our models could, therefore,

successfully distinguish the affinity of DHEAS.

There is some controversy in the literature as to whether unbound bilirubin is a

substrate for OATP1B1 and its inclusion in the meta-pharmacophore may necessitate

further evaluation. Although it was shown to be a substrate in both Hek-293 cells and

oocytes (Table 5.1), no transport was observed in HeLa and Hek-293 cells by Wang and

colleagues (158). We have therefore generated a meta-pharmacophore without bilirubin

and found that this pharmacophore no longer has three hydrophobic features and has less

optimal training set statistics due to the narrowing of the Km range. The mapping of

DHEAS to this modified meta-model is consistent with the initial model as it maps to both hydrophobic features and one H-bond acceptor feature (Figure 5.3C), suggesting that the hydrophobic feature removed after exclusion of bilirubin is not critical to the overall model. Our data are unlikely to end the controversy over whether or not bilirubin is a substrate for OATP1B1.

The approaches taken in this study indicate that relatively small amounts of literature data for drug transporters can be used to provide qualitative models of the key features involved in the ligand-protein interaction. By combining data from different experimental cell systems expressing the transporter, meta-models can be generated which comply with most criteria for robust pharmacophores. Namely these models have a sizeable difference between the null and total cost of the hypothesis and show a decrease in the average model correlation following scrambling of the molecule structures and the

Km values. External testing validated the predictive ability of the models. These meta-

128 pharmacophores are grossly similar to merged models of the multiple pharmacophores derived with individual datasets. This approach of merging individual models for the same transporter had previously been taken with P-gp to show the concordance of five inhibitor pharmacophores and the tight clustering of hydrophobic features at the extremities and central H-bonding features (30). The pharmacophores for OATP1B1 and

Oatp1a1 (Figure 5.3D-F) are distinctly different to those previously generated for P-gp in that the hydrophobic features are now centrally located pharmacophore features with the

H-bond acceptor features at the extremities.

In conclusion, we have shown for the first time that using limited data sets from different laboratories, cell types and species can be used successfully to derive robust computational pharmacophore models describing the key features for substrate interaction with rat Oatp1a1 and human OATP1B1 transporters. Our ultimate meta- pharmacophore approach combines data from different cell types and produces models which suggest a degree of similarity consistent with the biological data generated for each transporter to date (Figure 5.3D-F). We believe this meta-pharmacophore approach can be used with other transporters for which in vitro data are scattered across multiple experimental systems and species. Ultimately, this approach may aid in predicting potential drug-transporter or drug-drug interactions occurring at the key human transport proteins and overcome some of the variability due to the unique qualities of individual cell types and experimental systems used.

129

Molecule Km (µM) Cell type Reference BametR2 10 Oocyte (159) Bamet ud2 9.7 Oocyte (159) Bilirubin (unconjugated) 0.0076 Oocyte (160) Bilirubin monoglucuronide 0.1 Hek293 (155) Bisglucuronosyl bilirubin 0.28 Hek293 (155) Bromosulfophthalein (BSP) 0.14 Hek293 (155) Cholate 11.4 Hek293 (155) Demethylphalloin 17 Oocyte (161) (162) Demethylphalloin 39 Hek293 DHEAS 21.5 Hek293 (155) Enalapril 268 Hek293 Unpublished Pang and Keppler Estradiol 17β glucuronide 3.71 Hek293 (163) Estradiol 17β glucuronide 23.8 MDCK II (164) Estradiol 17β glucuronide 8.2 Hek293 (165) Estradiol 17β glucuronide 9.7 Oocyte (166) Estradiol 17β glucuronide 8.2 Hek293 (155) Estradiol 17β glucuronide 8.29 Hek293 (167) Estrone-3-sulfate 0.458 Hek293 (167) Estrone-3-sulfate 7 Hek293 (163) Estrone-3-sulfate 12.5 Hek293 (155) Estrone-3-sulfate 5.34 Oocyte (163) Pitavastatin 3 Hek293 (167) Pravastatin 13.7 Oocyte (166) Pravastatin 24.3 MDCK II (164) Pravastatin 35 Hek293 (168) Rifampin 1.5 HeLa (157) Rifampin 13 Oocyte (169) Rosuvastatin 8.5 Oocyte (170) Taurocholate 10 Hek293 (155) Taurocholate 13.6 Oocyte (171) Taurocholate 33.8 Hek293 (168) Thyroxine 3 Oocyte (171) Triidothyronine 2.7 Oocyte (171)

Table 5.1. Literature human Km data for OATP1B1

130

Molecule Km (µM) Cell type Reference Aldosterone 0.015 Oocyte (142) BSP 3 CHO-03 (172) BSP 3.3 HeLa (173) BSP 1.5 Oocyte (174) BSP 3.3 HeLa (175) BSP 1.5 Oocyte (176) BQ-123 600 Oocyte (177) Cholate 54 CHO-03 (172) Cortisol 13 Oocyte (142) CRC-220 57 CHO-03 (172) CRC 220 29.5 Oocyte (178) Deltorphin II 137 Oocyte (179) DHEAS 5 CHO-03 (172) DPDPE (enkephalin) 48 Oocyte (179) Enalapril 214 HeLa (153) Estradiol 17β glucuronide 2.58 LLC-PK1 (180) Estradiol 17β glucuronide 4 CHO-03 (172) Estradiol 17β glucuronide 11.2 LLC-PK1 (181) Estradiol 17β glucuronide 3.2 HeLa (182) Estradiol 17β glucuronide 20.4 Cos-7 (183) Estradiol 17β glucuronide 11 Cos-7 (184) Estradiol 17β glucuronide 3 Oocyte (142) Estrone-3-sulfate 12 CHO-03 (172) Estrone-3-sulfate 4.5 Oocyte (142) Fexofenadine 32 HeLa (185) Gadoxetate 3300 Oocyte (186) Glycocholate 54 CHO-03 (172) Hyodeoxycholic acid 17.5 HeLa (173) Leukotriene C4 0.27 Oocyte (145) Ouabain 3000 CHO-03 (172) Ouabain 1700 Oocyte (142) Ochratoxin A 29 CHO-03 (172) Ochratoxin A 16.6 Oocyte (187) Pravastatin 30 Hek-293 (168) S-dinitrophenyl glutathione 408 Oocyte (145) Taurocholate 32 CHO-03 (172) Taurocholate 50 Oocyte (174) *also known as sulfolithocholyltaurine

Continued

Table 5.2. Literature Km data for rat Oatp1a1

131

Table 5.2 continued

Taurocholate 91 Oocyte (188) Taurocholate 14 HeLa (144) Taurocholate 38 Oocyte (145) Taurocholate 19.4 HeLa (173) Taurocholate 27.2 HeLa (182) Taurolithocholic acid sulfate* 12.6 HeLa (189) Tauroursodeoxycholate 13 CHO-03 (172) Taurodeoxycholic acid 3.5 HeLa (173) Taurochenodeoxycholate 7 CHO-03 (172) Taurohyodeoxycholate 3.1 HeLa (173) Temocaprilat 46.7 Cos-7 (184)

132

mean human Km mean rat Km log mean human log mean rat Molecule (µM) (µM) Km (µM) Km (µM) BSP 0.14 2.52 -0.85 0.40 Cholate 11.4 54.0 1.06 1.73 DHEAS 21.5 5.00 1.33 0.70 Estradiol 17β glucuronide 7.91 8.80 0.90 0.94 Estrone-3-sulfate 6.32 8.25 0.80 0.92 Enalapril 268 214 2.43 2.33 Pravastatin 24.3 30 1.39 1.48 Taurocholate 19.1 38.8 1.28 1.59

Table 5.3. Summary of the molecules tested with human OATP1B1 and rat Oatp1a1

133

Dataset Correlation Total cost (model) Total cost (fixed) Total cost (null) OATP1B1 oocyte 0.97 62.52 61.27 60.57 trial_average 0.96 62.94 61.19 60.57 OATP1B1 Hek 0.91 70.11 63.85 67.51 trial_ average 0.76 69.77 58.43 67.51 OATP1B1 meta 0.92 88.00 81.48 103.84 trial_ave 0.84a 93.14 79.09 103.84 OATP1B1 meta 0.92 84.58 80.67 83.31 (no bilirubin) trial_ave 0.84 84.38 76.40 83.31 Oatp1a1 oocyte 0.92 74.35 63.53 107.66 trial_ average 0.87b 85.89 68.80 107.66 Oatp1a1 CHO 0.90 64.44 60.71 56.05 trial_ave 0.69 63.89 55.47 56.05 Oatp1a1 HeLa 0.95 53.33 52.62 37.08 trial_ average 0.90 52.20 50.88 37.08 Oatp1a1 meta 0.90 114.86 103.09 146.04 trial_ave 0.78 132.38 108.90 146.04 aaverage of 9 trials as one run failed = 0.75; baverage of 8 trials as 2 runs failed = 0.69

Table 5.4. Model building and scrambling (trial average) summary

134

OATP1B1-oocyte Feature aHBA HBA bHYD HYD HYD Weight 1.89847 1.89847 1.89847 1.89847 1.89847 Tolerance 1.60 2.20 1.60 2.20 1.60 1.60 1.60 Coordinates X 2.94 5.75 -3.93 -1.44 -5.90 -6.93 -1.27 Y -1.06 -1.62 6.72 7.74 -7.08 7.01 3.55 Z 0.09 0.97 0.77 -0.56 -0.72 1.97 1.73 do------> o------> eo o o OATP1B1-hek Feature HBA HBA HBA HYD Weight 2.42352 2.42352 2.42352 Tolerance 1.60 2.20 1.60 2.20 1.60 2.20 1.60 Coords X -3.57 -1.27 -4.65 -3.33 2.46 5.10 -3.14 Y 3.51 2.02 4.37 6.40 -4.91 -5.13 0.10 Z 6.03 7.25 2.58 0.81 -0.45 -1.90 -0.96 o------> o------> o------> O OATP1B1-meta Feature HBA HBA HYD HYD HYD Weight 1.97880 1.97880 1.97880 1.97880 1.97880 135 Tolerance 1.60 2.20 1.60 2.20 1.60 1.60 1.60 Coordinates X 4.27 6.98 -0.61 1.65 -6.25 -3.64 -5.24 Y 1.04 0.00 6.35 8.14 4.61 7.55 -11.09 Z 1.47 2.23 -3.08 -3.88 -2.22 -2.32 2.51 o------> o------> o o O aHBA = H-bond acceptor, bHYD = hydrophobic, cNegIonizable = negative ionizable feature, d0-----> = pharmacophore vector eo = pharmacophore point

Continued

Table 5.5. OATP pharmacophore coordinates

Table 5.5 continued

OATP1B1-meta-(no bilirubin) Feature HBA HBA HYD HYD Weight 2.14102 2.14102 2.14102 2.14102 Tolerance 1.60 2.20 1.60 2.20 1.60 1.60 Coordinates X -3.68 -2.79 4.18 4.02 -1.89 -2.13 Y 4.42 6.83 0.09 2.71 -2.87 -0.53 Z 2.93 1.36 1.72 0.26 -4.99 -0.83 o------> o------> o o Oatp1a1-oocyte Feature HBA HBA HYD HYD HYD Weight 2.44559 2.44559 2.44559 2.44559 2.44559 Tolerance 1.60 2.20 1.60 2.20 1.60 1.60 1.60 Coordinates X 2.71 2.89 -6.22 -8.38 -2.54 1.62 3.18 Y -0.74 -3.69 -0.59 -1.50 1.30 2.52 0.88

136 Z -3.73 -4.35 0.22 -1.67 0.00 -0.62 1.50 o------> o------> o o O Oatp1a1-CHO Feature HBA HYD HYD HYD Weight 1.66383 1.66383 1.66383 1.66383 Tolerance 1.60 2.20 1.60 1.60 1.60 Coordinates X -8.43 -10.12 -7.17 -4.19 -5.37 Y 0.83 0.46 -6.31 -3.23 -2.26 Z 0.94 -1.51 2.04 -0.19 3.14 o------> o o o

Continued

Table 5.5 continued

Oatp1a1-Hela Feature HBA HYD dNegIonizable Weight 2.00092 2.00092 2.00092 Tolerance 1.60 2.20 1.60 1.60 Coordinates X 2.51 3.70 3.67 3.80 Y -18.65 -17.83 -12.58 -9.03 Z 7.13 4.50 6.00 13.17 o------> O o Oatp1a1-meta-outlier (Leukotriene-c4, S-dinitrophenyl-glutathione) Feature HBA HBA HYD HYD HYD Weight 2.19247 2.19247 2.19247 2.19247 2.19247 Tolerance 1.60 2.20 1.60 2.20 1.60 1.60 1.60 Coordinates X -6.11 -8.16 3.28 3.24 3.18 -2.54 1.62 Y 0.27 -0.03 0.69 -2.04 0.48 1.70 2.52

137 Z -1.02 -3.19 -3.65 -4.89 1.90 0.00 0.18 o------> o------> o o o

Figure 5.1. OATP Pharmacophores generated from substrate data for human OATP1B1 expressed in (A) oocytes (showing bilirubin mapped to features) and (B) Hek cells (showing bilirubin monoglucuronide mapped to features), and rat Oatp1a1 expressed in (C) oocytes (showing aldosterone mapped to features), (D) CHO cells (showing BSP mapped to features), (E) HeLa cells (showing taurohyodeoxycholate mapped to features), (F) merged OATP1B1model using pharmacophores described in A and B, (G) meta analysis model using all cell type molecule data for human OATP1B1 (showing bilirubin mapped to features) and (H) merged Oatp1a1 model using pharmacophores described in C, D and E, with showing aldosterone mapped to features (I). Pharmacophore features include hydrophobes (cyan), negative ionizable (purple) and H-bond acceptors (green).

138

Figure 5.1

139

2 m (µM)

logK 1 ty i tiv c 0 d A te ic d -1 Pre

-2

-3 -3 -2 -1 0 1 2 3 4

Experimental Activity logK (µM) m

Figure 5.2. Test set predictions for the rat Oatp1a1-CHO pharmacophore. 15 molecules were used with Km data generated in other cell types.

140

Figure 5.3. Mapping of DHEAS to three pharmacophore models. Human OATP1B1 meta (A), rat Oatp1a1 meta (B), human OATP1B1 meta (no bilirubin) (C) model. Pharmacophore features include hydrophobes (cyan) and H-bond acceptors (green). Illustration of the interatomic distances within the pharmacophore models for human OATP1B1 meta (D), rat Oatp1a1 meta (E), and human OATP1B1 meta (no bilirubin) (F). Red spheres represent H-bond acceptor features whereas green spheres represent hydrophobic features. The labeled inter-feature distances are shown in Ångstroms. (G-N): Mapping of taurocholate (G,K), BSP (H, L), cholate (I, M) and estradiol-3-sulphate (J, N) to the oatp1a1 (G-J) and OATP1B1 (K-N) pharmacophores.

141

Figure 5.3.

142

Figure 5.4. Structural formulas of the compounds used in this study

143

CHAPTER 6

STRUCTURAL DETERMINANTS OF P-GLYCOPROTEIN-MEDIATED

TRANSPORT OF GLUCOCORTICOIDS

6.1 Introduction

Glucocorticoids, such as methylprednisolone, block the effects of inflammatory cytokines by entering the cell and binding to the intracellular glucocorticoid receptor

(GR). It was thought previously that glucocorticoids move freely into and out of the cell by simple diffusion only. However, Meijer and colleagues (190) demonstrated in vivo that penetration of dexamethasone into the mouse brain is limited by the presence of mdr1a. Karssen and co-workers (191) demonstrated that P-glycoprotein impairs the penetration of cortisol, but not corticosterone, into the brain of mice and humans. These studies suggest that subtle differences in chemical structure (e.g., absence of a 17α- hydroxyl group) may determine whether a glucocorticoid-like compound will be a substrate for P-glycoprotein. Nakayama and colleagues (192) showed that P-glycoprotein limited the oral absorption of glucocorticoids in a regional-dependent manner (duodenum

> jejunum > ileum) with a strong correlation to P-glycoprotein expression along the

144 longitudinal axis of the gastrointestinal tract. Interestingly, the absorption of sex steroids, such as progesterone, was unaltered by the expression of P-glycoprotein. These data corroborate the hypothesis that structural and/or physicochemical properties of individual glucocorticoids determine how efficiently they are transported by P-glycoprotein and that transport by P-glycoprotein plays an important role in determining their in vivo efficacy.

In vitro studies demonstrated the ability of P-glycoprotein to limit the transepithelial penetration and, hence, intracellular accumulation of glucocorticoids (193-

195). Early studies in this area suggested that glucocorticoid transport efficiency may be related to lipophilicity (196). However, Gruol and Bourgeois (194) demonstrated that various steroid analogs are transported by P-glycoprotein with varying efficiency, based on specific structural recognition sites that mediate P-glycoprotein binding and transport.

These authors (194) further showed that the number of hydroxyl groups, as well as their position of substitution, affects interaction with P-glycoprotein. For example, steroids lacking both 17- and 21-hydroxy groups are transported much less efficiently than prednisolone, a steroid containing 17- and 21-hydroxyl groups (Figure 6.1). Moreover, steroids devoid of hydroxyl groups (e.g., progesterone) are not transported by P- glycoprotein (195, 197). Thus, the presence of 11-, 17-, and 21-hydroxyl groups appears to be a critical determinant for transport efficiency of steroids. The current study was designed to investigate the structural elements of glucocorticoids responsible for interaction with P-glycoprotein.

Non-expressing (LLC-PK) and P-glycoprotein-expressing (L-MDR1) cells were used in transepithelial transport studies of clinically active glucocorticoids to characterize glucocorticoid structural features requisite for P-glycoprotein recognition and transport.

145

The resulting in vitro data were used to analyze and correlate structural features of glucocorticoids that contribute to their ability to be recognized and transported by P- glycoprotein. A comprehensive three-dimensional quantitative structure-activity relationship (3D-QSAR) approach was applied using various molecular descriptors.

These studies represent the first comprehensive in vitro and in silico examination of glucocorticoid transport by P-glycoprotein.

146

6.2 Methods

The experimental study was performed by Dr. Charles Yates and colleagues.

Chemicals

Cortisol (11α, 17β, 21-trihydroxypregn-4-ene-3,20-dione), cortisone (4-pregnen-

17, 21-diol-3, 11, 20-trione), betamethasone (9α-fluoro-16β-methylprednisolone), prednisone (17α-21-dihydroxy-1,4-pregnadiene-3,11,20-trione), prednisolone (11β, 17α,

21-trihydroxy-1,4-pregnadiene-3,20-dione), methylprednisolone (6α-methyl-11β, 17α,

21-trihydroxy-1,4-pregnadiene-3,20-dione), dexamethasone (9α-fluoro-16α- methylprednisolone), and verapamil were purchased from Sigma (St. Louis, MO, USA).

6α-hydroxy cortisol (4-pregnen-6α, 11β, 17, 21-tetrol-3, 20-dione) and 6β-hydroxy cortisol (4-pregnen-6β, 11β, 17, 21-tetrol-3, 20-dione) were purchased from Steraloids

(Newport, RI, USA). Calcein acetoxylmethyl ester (CAL-AM) was obtained from

Molecular Probes (Eugene, OR, USA). HPLC-grade acetonitrile was purchased from

Fisher Scientific (Fairlawn, NJ, USA). 3H-Dexamethasone was obtained from New

England Nuclear (Boston, MA, USA).

Cell Culture

LLC-PK, a pig kidney epithelium-derived cell line, and a variant cell line that stably expresses the product of the human P-glycoprotein gene (L-MDR1) were cultured in Media 199 supplemented with 10% fetal bovine serum, 2 mM L-glutamine, and 100

U/mL penicillin and streptomycin at 37 °C in a humidified incubator at 5% CO2. L-

147

MDR1 cells were cultured with 640 nM vincristine to maintain positive selection of

MDR1 over-expressing LLC-PK cells. LLC-PK and L-MDR1 cells were a generous gift

of Dr. Alfred Schinkel (The Netherlands Cancer Institute, Amsterdam, The Netherlands).

The L-MDR1 variant was characterized previously with respect to extent and specificity

of P-glycoprotein over-expression and monolayer growth (198).

The murine macrophage-like cell line J774.2 and an mdr1b over-expressing

variant (J774.2 mdr1b) were maintained in DMEM supplemented with 2 mM L-

glutamine, 100 U/mL penicillin, 100 µg/mL streptomycin, and 10% heat-inactivated

bovine serum. Heat inactivation was achieved by heating serum to 56°C for 30 minutes.

Cells were maintained at 37°C in a humidified incubator at 5% CO2. The J774.2 mdr1b variant was previously characterized with respect to extent and specificity of P- glycoprotein over-expression (199).

Transwell culture inserts used for transport experiments (3.0 µm pore size, 24.5 mm diameter, microporous polycarbonate membrane; Transwell™ 3414) and all other cell culture supplies were obtained from Costar (Cambridge, MA, USA) or Life

Technologies (Rockville, MD, USA).

Determination of MDR Phenotype in LLC-PK and L-MDR1 Cells

Calcein-AM (CAL-AM) is a lipophilic compound that freely passes the cellular membrane. Once inside the cell, the ester moiety is cleaved by ubiquitously expressed intracellular esterases yielding a highly fluorescent metabolite, calcein. CAL-AM is actively extruded by P-glycoprotein. In contrast, calcein, the hydrophilic free acid form, is not a substrate for P-glycoprotein (200). The kinetics of calcein formation can therefore

148

be used to assess functional expression of P-glycoprotein. LLC-PK and L-MDR1 cells (1

× 105) were plated in 96 well plates at a density of 5 ×105 cells/ml in Media 199. After overnight incubation at 37°C, media was aspirated and cells were washed with 37°C phosphate buffered saline (PBS, pH 7.4). CAL-AM (1 µM) was then added and calcein formation was monitored at 37°C using a Gemini SpectraMax plate reader (Molecular

Devices; Sunnyvale, CA, USA) set at excitation and emission wavelengths of 481nm and

536 nm, respectively. Differences in the rates of calcein formation (dCAL/dt) between

LLC-PK and L-MDR1 cells were used to assess expression of functional P-glycoprotein.

Parallel experiments were conducted in which CAL-AM (1 µM) was incubated in the absence of cells to account for calcein formation resulting from non-enzymatic ester cleavage. Additional control experiments were conducted to verify that differences in dCAL/dt were related to P-glycoprotein expression. Verapamil (100 µM), a known inhibitor of P-glycoprotein, was added to cells pre-loaded with CAL-AM (1 µM) and calcein formation was determined as described above. Verapamil is a P-glycoprotein substrate and inhibits transport function without interrupting the catalytic cycle (201).

Transepithelial Transport of Glucocorticoids

Transport experiments were conducted using LLC-PK (control) and L-MDR1 cells. LLC-PK cells form a polarized monolayer permitting bi-directional (basal to apical and apical to basal) assessment of drug flux. In L-MDR1 cells, P-glycoprotein preferentially sorts to the apical membrane enhancing basal to apical flux and inhibiting apical to basal flux. Cells were seeded on microporous polycarbonate membrane filters at

149

a cell density of 2 × 106 cells per well for LLC-PK and L-MDR1 cell lines. The cells were grown for 3 days with daily media changes. Two hours prior to transport, culture media was replaced with fresh media. For transepithelial transport measurements, the media on either the basal or apical side was replaced with media containing the glucocorticoid of interest (50 µM). The initial donor compartment glucocorticoid concentration (C0) was selected to achieve detectable receiver compartment

concentrations during the linear phase of transport. The cells were incubated at 37°C and

an aliquot (100.0 µl) of media from the opposite side of the monolayer (receiver

compartment) was withdrawn at 0.5, 1, and 1.5 hours.

Parallel experiments were conducted in which the flux for cortisol, prednisolone,

and methylprednisolone was determined in the absence of cells. These glucocorticoids

were chosen because they represent the range of partition coefficients found in our study.

Glucocorticoid (50 µM) was added to the apical well and an aliquot (100 µl) of media

from the basal well was withdrawn at 5, 10, and 15 minutes. Samples were stored at –

80°C until HPLC analysis. Each directional transport experiment was conducted in

triplicate. Monolayer integrity was verified by measuring transepithelial electrical

resistance (TEER) just prior (0 hr) to addition of drug and at the end of each transport

experiment (1.5 hr).

HPLC Assay

An aliquot (100.0 µl) from the receiver compartment was injected onto HPLC

(Alliance, Waters Associates; Milford, MA, USA). Standard curves for individual

150

glucocorticoids were constructed and validated over a range of 50 to 2000 ng/mL. The

HPLC system consisted of a pump, automated injector (Alliance, Waters Associates), and

a photo-diode array detector (model 996, Waters Associates). The stationary phase was a

Novapak C-18 column protected by a Novapak C-18 pre-column (Waters Associates).

The mobile phase consisted of 30% acetonitrile and 70% water and was delivered at a

flow rate of 0.9 ml/min. The eluent was monitored at 254 nm. Discrimination between

11-keto and 11-OH glucocorticoids was achieved using a normal phase HPLC assay

described elsewhere (202).

Calculation of Effective Permeability Coefficient (Peff)

Effective permeability coefficients (Peff) were calculated from concentration-time

profiles as measured in the receiver compartment according to Fick’s first law using the

following equation:

⎛ dQ⎞ V P = ⎜ ⎟ ⋅ eff ⎝ dt ⎠ A ⋅ C 0

where dQ/dt represents the appearance of glucocorticoid in the receiver chamber

(ng/ml/min), V is the volume of the receiver compartment (cm3), A is the cross-sectional area, and C0 is the initial donor concentration (ng/ml) at time = 0. The flux across the monolayer was determined by linear regression from individual glucocorticoid concentration versus time curves. Comparison of Peff in the basal (B) to apical (A) direction (Peff, B→A)) to Peff in the A to B direction (Peff, A→B)) in LLC-PK and L-MDR1 cells was used to assess transport efficiency (Teff = Peff, B→A / Peff, A→B).

151

Molecular Modeling

Structure sketching details have already been described in the introduction chapter.

DISCO

Cortisol was selected as reference compound for the data set. DISCO was initially run considering all the potential “feature” points. Additional runs with the specification of a minimum of two hydrophobic centers was also carried out. The resulting pharmacophore models were used to superimpose each set of the molecules. The resulting structural overlaps were analyzed by CoMFA.

FieldFit

In addition to structurally aligning the molecules using DISCO, the

SYBYL/MULTIFIT subroutine was used to implement the FieldFit technique. As

DISCO features are the reflection of the pharmacologic potency of compounds, it is reasonable to take these dummy atoms into account in the FieldFit procedure. DISCO features were assigned to the molecules before they were aligned. A control experiment in the absence of DISCO features was also performed to test the effectiveness of the new protocol.

CoMFA

CoMFA settings have already been described in the introduction chapter. The dependent variable (biological descriptor) used in these studies was the effective permeability (Teff) of glucocorticoid analogs in L-MDR1 cells. The experimental standard deviation was used as a weighting factor in the PLS analyses.

152

CoMSIA

CoMFA settings have already been described in the introduction chapter.

6.3 Results

Determination of MDR Phenotype in LLC-PK and L-MDR1 Cells

A fluorescence-based microtiter plate assay using CAL-AM was employed to

confirm the MDR phenotype of LLC-PK and L-MDR1 cells. Differences in the rate of

calcein formation (dCAL/dt) were used to assess P-glycoprotein expression. L-MDR1

cells were incubated in the presence of CAL-AM. Addition of verapamil (100 µM)

during CAL-AM treatment resulted in a significant increase in dCAL/dt (5.3 vs. 1.2

F*/min; p < 0.001) in L-MDR1 cells (Figure 6.2). As expected, in wild-type LLC-PK cells lacking P-glycoprotein, dCAL/dt was not significantly affected by verapamil (4.7 vs.

4.6 F*/min; NS), confirming that these cells do not express functional amounts of P- glycoprotein. In L-MDR1 cells, dCAL/dt in the presence of verapamil was approximately equal to dCAL/dt in the presence/absence of verapamil in LLC-PK cells, indicating that the differences in dCAL/dt are attributable to MDR1 expression in L-MDR1 cells. As

LLC-PK cells are also devoid of sister of P-glycoprotein (sPGP, also known as “bile salt export pump” (BSEP) or ABC-B11) (203) and multidrug resistance protein (MRP1,

ABC-C1) (204), they serve as an excellent model for in vitro studies of drug transport and were used for the remainder of our studies.

153

Transepithelial Transport of Glucocorticoids

Analysis of the transepithelial flux of cortisol in P-glycoprotein transfected LLC-

PK cells revealed that the effective permeability in the basolateral to apical direction (Peff,

B→A) of cortisol was approximately 3-4-fold greater than Peff, A→B (Figure 6.3, Table 6.1).

In our studies, we used the reciprocal of Peff, B→A over Peff, A→B , henceforth named Teff , as a measure of the increase in transport attributed to the P-glycoprotein-mediated component of the overall transepithelial flux (Table 6.1). As expected for P-glycoprotein substrates, control experiments using wild-type LLC-PK cells revealed no significant difference between Peff, B→A and Peff, A→B (i.e., Teff ≅ 1; Table 6.1). These data are

consistent with previous studies demonstrating P-glycoprotein transport of cortisol (195).

All other glucocorticoids tested were substrates for P-glycoprotein as evidenced by a Teff

value ranging from 1.8 to 26.6 in L-MDR1 cells (Table 6.1). Mean (± S.D.) Teff was 1.1 ±

0.17 for all glucocorticoids when their vectorial transport was evaluated in LLC-PK cells

(Table 6.1). Prednisolone, which contains two carbon-carbon double bonds in the A-ring, was transported approximately 4-fold more efficiently than cortisol, which contains only a single carbon-carbon double bond in the A-ring. The addition of a 6α-methyl substituent to prednisolone (i.e., methylprednisolone) resulted in an increase in Teff of

approximately 2-fold.

Conversely, the addition of 9α-fluoro and 16α/β-methyl groups to prednisolone

(i.e., betamethasone and dexamethasone) resulted in a reduction of Teff (± S.D.) from

12.5 ± 3.5 for prednisolone to 8.3 ± 0.94 and 3.7 ± 0.93 for betamethasone and

dexamethasone, respectively (Table 6.1).

154

In our study, we found no relationship between partition coefficient and Teff

(Figure 6.4A). However, a negative correlation (r2 = 0.84) between lipophilicity and Peff,

A→B was noted when glucocorticoid drug flux was evaluated in LLC-PK cells (Figure

6.4B). Significant hydrophobic interaction between glucocorticoids and the

polycarbonate membrane of the transwell plate is one possible mechanism to explain

these results. However, when Peff values for cortisol, prednisolone, and

methylprednisolone were determined in the absence of cells, there was no relationship

between partition coefficient and Peff (data not shown). Lipophilic compounds tend to diffuse rapidly across cellular membranes. However, compounds with log octanol water partition coefficients > 3.5 exhibit decreased transepithelial permeability coefficients with increasing lipophilicity (205). In addition, it has been proposed that transcellular movement of lipophilic compounds may be restricted owing to interactions with cytoplasmic proteins and partitioning into the intracellular membrane lipid bilayer (206,

207). Thus it is likely that the inverse relationship between permeability and lipophilicity was related to hydrophobic interactions of glucocorticoids with the plasma membrane, as well as cytosolic proteins.

DISCO

To gain an understanding of the binding mode of glucocorticoid analogs to P- glycoprotein, we used a computational approach to model in vitro affinity data. DISCO

generated pharmacophore models of the position of important ligand features in three-

dimensional space that may ultimately relate to features within P-glycoprotein. These

models were derived using multiple conformations of each individual ligand alongside

155

the experimental inhibition data. The result is a computational model that can be used to

predict the affinity of glucocorticoids to P-glycoprotein and guide the design of novel

glucocorticoid receptor ligands.

A DISCO generated pharmacophore (Figure 6.5) consisted of 4 hydrophobes,

defined as a contiguous set of surface accessible atoms not adjacent to any charge; one H-

bond donor sites and a H-bond acceptor atom. This model has an rms·fit of 0.236 and a

union value of 480.6 Å3 (versus 284.9 for the null hypothesis).

CoMFA

The model depicted in Figure 6.6 has a cross-validated r2 (q2) of 0.43 with 4

components and a traditional (non cross-validated) r2 of predicted vs. actual Teff of 1.00, as illustrated in Fig 7A. The residual value of the predicted Teff for each individual compound does not exceed 0.25 of its experimental value (Figure 6.7B), indicating an internally consistent model. The press was 8.62 and s was 0.24. The relative contributions of the steric and electrostatic fields were 0.45 and 0.55, respectively.

CoMSIA

The CoMSIA model was derived from identical structural alignments as described for CoMFA. Surprisingly, H-bonding and hydrophobic descriptor fields did not contribute significantly to the overall quality of the initial LOO cross-validated models, despite the prevalence of these descriptors in the pharmacophore model. The optimal number of components for a model that included both electrostatic and steric fields was 5 with a q2 of 0.43 (press, 9.90; s 0.44). The relative contribution of the steric and

156

electrostatic fields was 0.88 and 0.12, respectively (model not shown). The best model

was derived when steric fields alone were considered with 5 components at a q2 of 0.63, resulting in a conventional r2 of 0.998 (press, 7.96, s 0.52). This model (Figure 6.6B) corroborates the previously identified steric CoMFA fields (Figure 6.6A), but displays more detail in the sterically favored and disfavored areas.

A better ability to visualize and interpret the obtained correlations in terms of field contributions is the major advantage of CoMSIA compared to standard CoMFA. In

CoMFA, the steric fields and electrostatic fields are treated simultaneously, and individual field contributions cannot be revealed. Essentially, the graphs represent isocontours of the obtained coefficients from PLS, indicating those lattice points where a particular property significantly contributes, and thus explains the variation in biological activity data. For example, the CoMSIA model indicates a sterically unfavorable region

(yellow) over 1-hydrogen of the A ring (Figure 6.1), that is not revealed in the CoMFA model, which would explain the higher transport efficiency of prednisolone that contains a 1,4-diene moiety in this position.

157

6.4 Discussion

Over the past decade, there has been an increasing body of evidence to suggest

that glucocorticoid analogs are substrates for the efflux transporter P-glycoprotein (190,

191, 208). Furthermore, subtle differences in molecular structure appear to strongly affect

transporter affinity (192, 195, 197). In the current study, we sought to test the hypothesis

that glucocorticoid receptor agonists are P-glycoprotein substrates and to directly

correlate P-glycoprotein affinity of glucocorticoid analogs to structural features using in

vitro transporter assays and molecular modeling approaches.

In the absence of a high-resolution structure for human P-glycoprotein, its

structural requirements can be inferred merely by indirect modeling techniques, using

variations in molecular structure associated with a change in P-glycoprotein affinity to

derive putative transporter-interaction points. However, the complexity of transport by

and modulation of P-glycoprotein has been a limiting factor in successful model

development. Recently, a study by Ekins and co-workers (30) successfully integrated a

structurally diverse data set into a single pharmacophore model for P-glycoprotein

affinity. The current study was initiated to expand the P-glycoprotein substrate data set

further by determining the structure-activity relationship of a series of glucocorticoid

analogs.

In vitro transport studies using transfected cell lines demonstrated that cortisol

and prednisolone are better substrates for P-glycoprotein than cortisone and prednisone,

respectively, as evidenced by their relative Teff values (Table 6.1). From a structural point of view, our data demonstrate enhanced transport of 11-hydroxyl substituted glucocorticoids as compared to 11-keto glucocorticoids and support prior suggestions by

158

Gruol and Bourgeois (194). Interestingly, prednisone and cortisone, which lack an 11- hydroxyl substituent, are substrates for P-glycoprotein. These data suggest that 11- hydroxyl substitution is not essential for recognition and transport by P-glycoprotein.

This conclusion is further supported by control experiments conducted in our laboratory using a normal phase HPLC assay (data not presented) showing that inter-conversion of

11-hydroxyl and 11-keto steroids (e.g., cortisone conversion to cortisol) did not occur in these cells.

P-glycoprotein recognition and transport of glucocorticoids is affected by other steroid ring substituents. 9α-Fluoro and 16α/β-methyl substitution (i.e., betamethasone and dexamethasone) reduced Teff. Interestingly, 16α-methyl substitution resulted in a greater reduction in Teff compared to 16β-methyl substitution (i.e., dexamethasone Teff was 2-fold lower than betamethasone). These data suggest that P-glycoprotein recognition and transport is sterically hindered by α-orientation of the methyl substitution.

Substitution and conformation of the A-ring also affected glucocorticoid recognition and transport. 6α-Methyl substitution (i.e., methylprednisolone) dramatically increased transport whereas 6α/β-hydroxyl substitution (i.e., 6α/β-hydroxy cortisol) reduced transport. A hydrophobic pocket within the steroid recognition site of P-glycoprotein would explain enhanced transport of methylprednisolone and reduced transport of 6α/β- hydroxy cortisol. Lastly, A-ring planarity affects Teff as prednisolone was transported approximately 4-fold more efficiently than was cortisol. Taken together, these data demonstrate that structural recognition plays an important role in determining the efficiency with which P-glycoprotein recognizes and transports glucocorticoids.

159

To further explore the structural characteristics of glucocorticoid affinity to P- glycoprotein, we employed a distance constraint pharmacophore building technique as featured in DISCO. Models were built by assessing multiple conformations of each substrate alongside their experimental transport data. DISCO generated 4 unique pharmacophores that, upon detailed inspection, were essentially similar in that the positioning of H-bond donor sites and H-bond acceptor atoms were interchanged. The consensus model presented in Figure 6.5 features four hydrophobic centers, two H-bond donor groups within P-glycoprotein and a H-bond acceptor atom within the glucocorticoid backbone; interatomic distances between groups are presented in Table

6.2. Our model for glucocorticoid analogs corresponds very well with the vinblastine transport model recently described by Ekins and colleagues (70), who list four ring aromatic features (hydrophobic centers) and two hydrogen acceptor sites. A common limitation to pharmacophore models is their failure to report on steric and electrostatic functionalities that drive short and long range ligand-protein interactions, respectively.

Therefore, we extended our model by correlating variability in transport efficiency to variations in molecular structure by implementing 3D-QSAR techniques. Using the molecular overlap from the previously developed pharmacophore model, CoMFA and

COMSIA were implemented to inspect the steric, electrostatic and H-bonding fields surrounding the molecules. The resulting models (visualized in Figs. 6A and 6B, respectively) illustrate areas of the molecules where alterations contribute significantly to

P-glycoprotein transport. It should be noted that the model was developed using P- glycoprotein transport data and therefore, the model reveals molecular regions that contribute to rendering a molecule a P-glycoprotein substrate. For example, both CoMFA

160 and CoMSIA models suggest less steric bulk within the vicinity of C-1 and C-11 may enhance P-glycoprotein-mediated transport. Regarding substitutions on the steroid D-ring, the CoMFA model suggests that electropositive moieties surrounding the C20,21 positions (with C17, S stereochemistry) may contribute to P-gp affinity. In turn, these observations could be used to specifically design compounds with lower P-glycoprotein affinity and, thus, lead to higher intracellular concentrations.

In vivo studies established the importance of P-glycoprotein in determining tissue penetration of glucocorticoids. For example, Schinkel and colleagues reported increased brain accumulation of dexamethasone, digoxin, and cyclosporin A in mdr1a knockout mice compared to wild-type mice (198). Interestingly, Saitoh et al. found that intestinal

P-glycoprotein limits the oral absorption of methylprednisolone, but not of cortisol (209).

The ability of P-glycoprotein to limit glucocorticoid tissue bioavailability might have important clinical implications as well. Koszdin et al. found that P-glycoprotein is responsible for limiting spinal cord bioavailability of methylprednisolone after intravenous administration (210). Thus, limited spinal cord bioavailability of methylprednisolone secondary to P-glycoprotein expression may explain the need for high-level systemic exposure of methylprednisolone in patients with acute spinal cord injury.

As previously suggested, circumvention of P-glycoprotein’s effects on steroid activity may be accomplished through development of steroids that do not serve as P- glycoprotein substrates (194). Our group is working on structure-based design of pharmacologically active glucocorticoids that are not transported by P-glycoprotein. In conclusion, our study provides an analysis of glucocorticoid structural features required

161 for recognition by the efflux transporter P-glycoprotein. These data are significant for several reasons. First, these data identify new structural features important in recognition/transport of glucocorticoids (e.g., 6α-substitution). Second, these data provide the framework for studies designed to further elucidate the glucocorticoid recognition site(s) of P-glycoprotein. Third, these data illustrate the potential importance of glucocorticoid selection when targeting tissues with significant P-glycoprotein expression. Overall, our study contributes to the emerging field of P-glycoprotein pharmacophore and 3D-QSAR development that can guide structure-based design of compounds that either target or avoid P-glycoprotein.

162

K .15 .13 .10 P .25 .29 .34 .28 .49 .36 ± ± ± - ± ± ± ± ± ± C ff S.D.) L e ± L 0.89 0.80 0.92 1.3 T ( 1.1 1.2 1.0 1.2 1.1

1 c

7.0 3.5 R .93 1.2 .94 .44 .93 .69 .44 ± ± /sec); D ± ± ± ± ± ± ± M ff S.D.) - e ± cm L 12.5 26.6 2.3 3.8 8.3 3.6 3.7 T ( 1.8 2.4 -2 b

A 3.5 2.3 0.5 .07 1.1 .67 2.0 .59 .15 → ± ± ± ± ± ± ± ± ± B S.D.) ff, e cient in the apical (A) to basal (B) ± i 11.6 14.8 12.2 0.30 8.0 7.9 7.3 5.8 P ( 1.2 f (35) f

A .07 .06 R .18 2.0 1.4 .05 .04 1.4 .35 → ± ± D cient; ± ± ± ± ± ± ± B i f M S.D.) ff, - e ± eability coef L 0.19 0.52 1.0 5.3 2.2 2.0 0.2 4.2 1.0 P ( ined in L-MDR1 cells (x10 , perm

B → Determ a

A A .05 .12 ff, 1.3 1.9 0.7 1.5 1.2 1.3 1.2 → ± ± e ± ± ± ± ± ± ± B S.D.) ff, e ± 0.34 0.80 4.4 7.4 5.9 4.6 4.0 9.1 5.3 P ( (34); PC, partition coef e

K B .07 .18 P → 1.2 0.6 0.5 0.6 0.8 1.0 1.8 ± ± - ± ± ± ± ± ± ± A C S.D.) ff, L e ± L P 3.7 9.0 6.4 4.5 3.7 8.0 6.6 ( 0.30 0.76 ined in LLC-PK cells; P cient in the B to A direction; i f e e e e e f e 7 8 7 7 4 8 2 ...... 7 8 7 0 1 5 6 determ ined in LLC-PK cells; ined in L-MDR1 cells 8 6 2 7 4 2 3 PC

B → A ff, e ters of glucocorticoid transepithelial transport ) determ ) determ e eability coef B B → →

A e and P A n A ff, ff, o e e , perm

→

A e d e B sol

n / P / P n → e oi o ff, A A o

s e n c B s e i a o → →

a ff, n h e e h t B o

B t predni n e sol s l e ff, ff, ol o y e e s s m i i h a t -OH cortisol -OH cortisol ucocort e α β Gl bet cort Predni cort 6 Predni m 6 dexam = (P = (P f f ef ef Average of P T T Table 6.1. Kinetic param a c d direction; P

163

Hy_1a Hy_2 Hy_3 Hy_4 DS_1 DS_2 AA_1 Hy_2 2.34 Hy_3 4.08 2.56 Hy_4 6.08 4.17 2.26 DS_1 4.77 6.14 8.38 9.85 DS_2 6.86 6.5 4.68 5.9 11.6 AA_1 4.06 3.68 2.55 4.57 8.83 2.9 AA_2 6.75 5.4 3.3 2.17 9.85 6.05 5.23 aAA, acceptor atom, DS, donor site; Hy, hydrophobic center; functional group names correspond to those depicted in Figure 6.5.

164

Table 6.2. Relative intramolecular distances between pharmacophoric feature points (Å)

O 21 OH 18 20 OH R 1 12 17 19 11 13 H 16 14 9 15 1 2 10 8 H H 3 5 7 4 6 O

–R1 –R2 Cortisol –OH –H Cortisone =O –H 6α-OH-Cortisol –OH OH 6β-OH-Cortisol –OH OH

O OH

OH R1

H R4

R3 H

–R1 –R2 –R3 –R4 Prednisolone –OH –H –H –H 6-Methylprednisolone –OH –CH3 –H –H Prednisone =O –H –H –H Betamethasone –OH –H –F CH3 Dexamethasone –OH –H –F CH3

Figure 6.1. Chemical structures of glucocortocoid receptor substrates. Backbone numbering according to IUPAC guidelines. Steroid cyclic rings are named as follows: ring A comprises C1-5,10; ring B, C5-C10; ring C, C8,9,11-14; ring D, C13-17.

165

Figure 6.2. Calcein formation (dCAL/dt) in L-MDR1 and LLC-PK cells. dCAL/dt in the presence (closed circles) or absence (open circles) of 100 µM verapamil in LLC-PK cells. dCAL/dt in the presence (closed squares) or absence (open squares) of 100 µM verapamil in L-MDR1 cells. Data represent mean (± S.D.) of three independent experiments.

166

Figure 6.3. Transepithelial transport of cortisol. Cortisol flux was measured from the apical to basolateral direction (A-B) (open symbols) and basolateral to apical direction (B-A) (closed symbols) in LLC-PK (circles) and L-MDR1 (squares) monolayers on microporous membrane filters. Each time point represents the mean (± S.D.) of three independent experiments.

167

Figure 6.4. Teff and Peff versus PC. Transport efficiency (Teff) versus partition coefficient (PC) as determined in L-MDR1 cells (Panel A). Data represent the mean (± S.D.) of three independent experiments. Plot of permeability (Peff) versus partition coefficient (PC) as determined in LLC-PK cells (Panel B). Reported Peff values were calculated by averaging Peff values in the apical to basal and basal to apical directions. Data represent the mean of three independent experiments.

168

Figure 6.5. Pharmacophore model derived from P-glycoprotein-mediated transport efficiency in L-MDR1 cells. The wire frame representation of all individual glucocorticoid analogs used in this study is shown in their overlapping conformations. Pharmacophore feature points (shown by red dummy atoms) are annotated as follows: AA, H-bond acceptor atom, DS, H-bond donor site; H, hydrophobic center.

169

Figure 6.6. CoMFA and COMSIA models for glucocorticoid analogs. The CoMFA isocontour map (a) of the electrostatic and steric contributions to logTeff around cortisol shows red contours to indicate regions where negative charge is favorable, whereas blue contours indicate regions where positive charge is favorable. Yellow contours represent regions where steric bulk is unfavorable; green contours indicate areas of the molecules where steric bulk is favorable. Carbon atoms are shown in white, hydrogen atoms in cyan, lone pairs in purple, and oxygen atoms in red. DISCO-derived pharmacophoric feature points are shown as dummy atoms in green. The COMSIA contours (b) illustrate areas where steric bulk is desirable (green) and molecular regions where steric moieties negatively affect logTeff (yellow).

170

1.6 A

1.4

1.2 f ef

0.8 Predicted LogT 0.6

0.4

0.2 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Actual LogT eff

10 B

Ligand # 4

0 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 Residual Value (LogT ) eff

Figure 6.7. Experimental versus predicted logTeff plot of the final CoMFA model for individual glucocorticoid analogs (top). The linear regression line has a slope of 1.0 and an r2 of 1.00. The individual residuals for all compounds are depicted in the bottom graph.

171

CHAPTER 7

USING PHARMACOPHORES TO RAPIDLY IDENTIFY MOLECULES WITH

AFFINITY FOR P-GLYCOPROTEIN

7.1 Introduction

Multidrug resistance (MDR) has become a major obstacle in the treatment of cancer due to over-expression of MDR pumps including P-glycoprotein (P-gp, ABCB1).

At the same time the transporter has a major role in determining the absorption of some drugs. P-gp is an ATP-dependent efflux transporter that transports a diverse range of structurally and functionally unrelated substrates across the plasma membrane. This transporter hence acts as a barrier limiting exposure to a diverse range of structurally and functionally unrelated substrates. To account for the observed broad substrate specificities for P-gp, the presence of multiple drug binding sites which can accommodate multiple molecules simultaneously have been proposed (211-215). P-gp is mainly expressed in the canalicular domain of hepatocytes, brush border of proximal

172 tubule cells and capillary endothelial cells in the central nervous system (CNS). This results in reduced oral drug absorption and enhanced renal and biliary excretion of substrate drugs. Despite its overall significance, P-gp is poorly characterized at the atomic level due to difficulties related to membrane protein crystallization, however combined photaffinity (216) and protein homology models have been developed (19,

217). In addition, a number of quantitative structure activity relationship (QSAR) and pharmacophore models have been generated (108, 218). A pharmacophore model describes the spatial arrangement of structural features responsible for activity toward a protein. Computational pharmacophores have been generated to predict the inhibition of

P-gp from in vitro data for several cell systems, including: structurally diverse inhibitors of digoxin transport in Caco-2 cells; vinblastine and calcein accumulation in P-gp expressing LLC-PK1 (L-MDR1) cells and vinblastine binding in vesicles derived from

CEM/VLB100 cells (30, 70) Most of these latter models correctly rank-order the data from the other probes, indicating partial overlap for the binding sites probed by digoxin and vinblastine. By merging all the P-gp inhibitor pharmacophores, common areas of identical chemical features such as hydrophobes, H-bond acceptors and ring aromatic features were apparent (70). Additionally, a common features HIPHOP alignment of the

P-gp substrates verapamil and digoxin produced a pharmacophore to which vinblastine partially aligned indicating possible affinity for a similar or identical binding site(s) within P-gp. The substrate and inhibitor pharmacophores overlapped to a considerable extent (70). Further pharmacophore-based approaches (219) using GASP alignments to vinblastine and to rhodamine123 have been carried out to understand the verapamil binding site (219). These studies have revealed similar pharmacophore requirements as

173 proposed earlier as multiple hydrophobic and H-bonding interactions were described

(219). The superimposition of a small number of P-gp ligands with SYBYL and

MOLCAD was also undertaken by Garrigues et al. to generate two pharmacophores

(220). The resulting models were validated with two additional compounds. The previous chapter also described a P-gp 3D-QSAR model and a pharmacophore model generated based on the transport profile of a set of glucocorticoids compounds.

However, the application of QSAR models and pharmacophores for database screening and identification of potential P-gp substrates in drug discovery are still limited.

A recent pharmacophore based search of the Derwent World Drug Index identified 28 diverse P-gp inhibitors (221) and represents one of the first approaches to using such models to prospectively identify molecules as inhibitors of this transporter. In the present study we have undertaken a comparison of different P-gp pharmacophore models (based on substrates and inhibitors) and applied them to search databases. Two P-gp digoxin inhibition models and one P-gp substrate model were generated and evaluated by searching a database of known P-gp substrates and non-substrates (222) to determine the optimal mode of searching. To further evaluate how the models performed in searching even larger commercial databases for known P-gp substrates we screened nearly 60,000 molecules with all three pharmacophores. The pharmacophores were also used to search a database of widely prescribed drugs in U.S. (98) for potential P-gp substrates.

Molecules were identified for further in vitro testing. These and other tests with recently published P-gp inhibitors demonstrate how pharmacophores for this and other transporters are of value for identifying potential molecules with affinity (34).

174

7.2 Materials and Methods

P-gp pharmacophore development with Catalyst

The detailed hypothesis generation settings have already been described in the previous introduction chapter. Briefly, the three models were constructed to generate the hypotheses. The previously published P-gp substrate HIPHOP model (70) based on verapamil and digoxin was regenerated. Since the substrate HIPHOP model is limited to identifying only large molecules a second model, the previously described P-gp digoxin inhibition hypogen model (30) (Inhibitor model 1) was also regenerated. To further update this latter model, 6 previously described test compounds with activity data were used (FK506A, PSC833, ritonavir, erythromycin, cyclosporine and CP99542) and incorporated (30) resulting in a second more comprehensive digoxin inhibition hypogen model (Inhibitor model 2) with 33 compounds.

Scrambling as described in the previous introduction chapter has been performed with inhibitor model 2.

Validation of the Catalyst models

The various models were tested with molecules not included in the initial training sets. These test set molecules were fit by the fast-fit algorithm to the respective Catalyst models in order to predict a value as previously described for CYPs (151). Fast fit refers to the method of finding the optimum fit of the substrate to the hypothesis among all the conformers of the molecule without performing an energy minimization on the conformers of the molecule.

175

Database searches

The three P-gp pharmacophore models were used to search different sized databases to provide a means of evaluation of their utility. The first database (P-gp database) consists of 189 known P-gp substrates and non-substrates (222) and was used to evaluate the performance of each model as well as the different searching methods.

Both the Fast search method and Best search method were applied during searching of this database. The second in-house database (SCUT) has already been described in chapter 3. The pharmacophore was then used with the fast-flexible approach to identify potential substrates in this database for P-gp. The third database used was the result of merging the commercially available drug-like molecules in Maybridge with 105 known

P-gp substrates from the first P-gp database (merged Maybridge database, with 59779 compounds). This database was used to further evaluate the performance of the different pharmacophore models.

Hit list analysis

The search results from P-gp database and merged Maybridge database were analyzed using a number of metrics described previously (223) (Figure 7.1). These metrics included the enrichment (E) which indicates the ratio of yield of active compounds in the hit list relative to the yield of active compounds in the database.

Secondly the yield (%Y) was determined which is the percentage of known active compounds in the hit list. Thirdly the coverage (%A) was measured which is the percentage of known active compounds retrieved from the database. Finally the Güner-

176

Henry score was derived which is a combination of yield and coverage with the

correction of hit list size (223). To be more accurate, the correction is the ratio of the

retrieved false positive hits over the inactive compounds in the database. Should the

whole database be retrieved as a hit list, this correction will add a penalty since this won't

be reflected by either yield or coverage. Small but selective hit lists (high %Y but low

%A) are very useful as this will represent fewer false positive hits in the hit list.

Therefore the yield is more important than coverage and is given a weight of 0.75, while

coverage has a weight of 0.25 in this formula (Figure 7.1).

7.3 Results

In the present study two literature derived pharmacophores for P-gp inhibitors and

substrates (Inhibitor model 1 and substrate model, respectively) have been previously

described. A third Catalyst P-gp pharmacophore model (Inhibitor model 2) was

developed with the 33 molecules (activity range 0.024 – 100 µM) and consisted of 4 hydrophobic (HYD) features and 1 H-bond acceptor (HBA) features (Table 7.1, Figure

7.2). The observed versus predicted correlation resulted in an r value of 0.87 and excellent models statistics for the difference between the total cost and null cost (Table

7.2). Following scrambling of this hypothesis and the activity data ten times, the average training correlation decreased to 0.56 and the difference between the total cost and null cost narrowed considerably (Table 7.2).

The three pharmacophores were then used to search the P-gp database. The results were analyzed using several metrics (defined in Figure 7.1) (223) and are listed in Table

7.3. The generally higher GH scores for the P-gp database using the fast search compared

177 to the best search in two out of three cases indicates the usefulness of fast search approach. This is advantageous as it is less expensive for use with larger databases. The substrate pharmacophore had the highest GH scores out of all the models and this is likely due to the higher enrichment and yields observed with this very specific pharmacophore. The Fast search approach was subsequently used to search the

Maybridge database. Once again the substrate pharmacophore model search result has a higher yield and lower coverage due to its more restrictive nature (more features that are distant) when compared with the inhibitor models.

The results of the database searches can be visualized to assess the overlap in the molecules identified by the various pahrmacophores. Figure 7.3a represents a Venn diagram showing the overlaps of all retrieved known P-gp substrates in the P-gp database with the three pharmacophore models. It is clear there is a high degree of overlap between the inhibitor models while the substrate model identifies a smaller subset of molecules shared between all three models. The average properties of each set of retrieved hits have clearly shown that being the largest pharmacophore model, the substrate model pulled out relatively large compounds characterized by the higher number of H-bond accepters, donors and larger molecular weight. The compounds that were retrieved by inhibitor model 1 and inhibitor model 2 have similar values for these properties. After demonstrating that all of the pharmacophores can be used to retrieve known P-gp substrates from the initial database, they were applied to the SCUT database for both further validation and to identify additional drugs as potential P-gp substrates.

All of the returned molecules from each pharmacophore based database search were subjected to a thorough literature analysis to determine if they had already been

178 experimentally confirmed as P-gp substrates. 40 drugs were retrieved by the Inhibitor model 1, out of these 25 were known substrates or inhibitors of P-gp based on the literature. For the Inhibitor model 2, 68 drugs were retrieved and 33 verified as known P- gp substrates. For the substrate model, 4 of the 6 returned drugs were known P-gp substrates. The Venn diagram illustrating the overlap of retrieved positive hits from the

SCUT database is shown in Figure 7.3b. As we do not have an estimate of how many known P-gp substrates are in the SCUT database in total, no statistical analysis was performed on the search results as previously performed on the other two databases.

A recent study (Szakacs et al., 2004) identified 18 compounds to be potential P-gp substrates. Some of these molecules had been experimentally verified as P-gp substrates.

We have used these molecules as a further test set for all pharmacophores (Table 7.5).

Eight of the molecules are assigned as experimentally confirmed substrates. The substrate model selects 3 of these molecules correctly and additionally assigns another 3 molecules as substrates. The inhibitor models indicated that most molecules had a high affinity for

P-gp (Table 7.5).

As our search of the SCUT database with Inhibitor model 2 retrieved a large number of molecules out of which approximately half were known P-gp ligands, we selected several of the remaining compounds for in vitro testing based there being no reported P-gp interaction in the literature and.on their commercial availability (ongoing project in cooperation with Praveen Bahadduri). The molecules are shown mapped to the pharmacophore (Figure 7.2). In addition all these molecules were scored with the other pharmacophores and the fit values were generated for comparison with the experimentally derived data.

179

7.4 Discussion

To date there have been numerous published P-gp models for substrates and inhibitors of P-gp (27, 29, 30, 70, 218, 222, 224) but there have been few applications of them to identify novel ligands that had not previously been identified (221). Recently we have used a pharmacophore for the uptake transporter hPepT1 to search a database and aid in the selection of molecules for in vitro testing. Subsequently we have been able to identify 3 molecules with affinity to this transporter that had not previously been described (34). In this study three distinct P-gp pharmacophore models were used for database screening as both a means of validation and to assist in selection of molecules for in vitro testing. All three models and especially the P-gp substrate HIPHOP model proved useful in enriching and identifying the number of P-gp substrates from non- substrates based on the calculated GH scores. Even though two inhibitor pharmacophore models were used, they were capable of selecting a considerable number of P-gp substrates as indicated by the GH scores, which were slightly lower than for the substrate model. After seeding a very large database of nearly 60,000 primarily drug-like molecules with a set of known P-gp substrates we were again able to demonstrate the validity of all 3 models in returning the known substrates. In this case the GH scores for the merged Maybridge database were considerably lower since the analysis assumes that the merged P-gp substrates are the only P-gp substrates and none of the Maybridge compounds are substrates. In the case of a large proportion of the Maybridge compounds being P-gp substrates, the tabulated results are perhaps therefore not unreasonable. In future it might be possible to test some of the Maybridge compounds indicated as P-gp substrates to verify these predictions.

180

Another recent study (225) applied RT-PCR to measure the correlation of ABC

transporter mRNA levels with drug resistance. The authors identified 18 compounds to

be potential P-gp substrates based on the activities being statistically significant in

negatively correlating with the expression of this transporter. Three of the 18 identified

hits had been previously experimentally verified as P-gp substrates and the authors

applied the MTT assay to verify that a further 6 available compounds were also P-gp

substrates. This set of 18 molecules retrieved from the NCI database were converted into

a Catalyst database with multiple conformations and used as a test set. Table 7.5 lists the

results of mapping all 18 compounds to different P-gp pharmacophore models. The

substrate pharmacophore was able to retrieve 3 of the 9 known substrates and also

suggested an additional 3 molecules for which there is presently no experimental

verification.

The P-gp pharmacophore models were also used to screen a database of known

drugs that are widely prescribed. In this case not only were we able to select some known

P-gp ligands but also 9 molecules were obtained for in vitro testing using the calcein

accumulation assay (Figure 7.4). This list included one molecule, repaglinide which had

previously been identified by a pharmacophore database search for hPepT1 (34). These

molecules were scored with all of the pharmacophores to enable a comparison of

predicted and experimental activity values. As a comparison the fast fit of verapamil to

the substrate pharmacophore has a score of 4.54 and LY335979 has a score and IC50 value for Inhibitor model 1 (3.29, 1.5µM) and Inhibitor model 2 (5.72, 23µM). The scores and predicted IC50 value for the molecules selected are similar to these standard substrate and inhibitor molecules.

181

The pharmacophore based screening outlined in this study using different databases could represent a method to either, filter out P-gp ligands, identify potential thereapeutic P-gp inhibitors as well as gauge the substrate coverage of various commercial databases prior to purchase. This would be particularly useful to avoid certain compounds or efficiently select those for in vitro screening for lead identification of potential P-gp inhibitors for therapeutic use. In the current study we have demonstrated the use of the GH score (223) as a metric for assessing the efficacy of a pharmacophore results. We have also demonstrated how models generated with limited numbers of small molecules can be used to efficiently identify other known ligands not included in the model as well as suggest new molecules for testing. These results compare favourably and extend the recent P-gp inhibitor database search approach (221) and complement our results with hPepT1 (34) as well as the Catalyst search studies performed by others (226-

229) to discover ligands for other proteins. The combination of an iterative in silico and in vitro approaches represent a means to more rapidly suggest and then verify known drugs or novel molecules from vendor libraries as ligands for these proteins.

182

Feature aHBA bHYD HYD HYD HYD Weight 1.93 1.93 1.93 1.93 1.93 Tolerance 1.60 2.20 1.60 1.60 1.60 1.60 Coordinates X 5.67 7.83 3.66 -5.12 6.44 6.64 Y 2.86 4.94 -1.50 -1.44 0.38 0.34 Z 0.92 1.10 4.67 1.82 -0.72 3.15 do------> eo o o o aHBA = H-bond acceptor, bHYD = hydrophobic, d0-----> = pharmacophore vector eo = pharmacophore point

Table 7.1. Coordinates for inhibitor pharmacophore 2. HBA = H-bond acceptor, HYD = hydrophobic features.

183

Correlation Total cost (model) Total cost (fixed) Total cost (null) Inhibitor Pharmacophore 2 0.87 151.8 130.7 198.8 Scrambling average 0.56 186.3 127.2 198.8

Table 7.2. Model statistics for the inhibitor model 1. The average results after scrambling 10 times is also listed.

184

Model and search type Enrichment (E) Yield (Y%) Coverage (A%) GH-Score Inhibitor model 1 fast 1.4 78 27 0.59 Inhibitor model 1 best 1.3 74 30 0.55 Inhibitor model 2 fast 1.31 72 30 0.53 Inhibitor model 2 best 1.32 73 45 0.53 Substrate model fast 1.8 100 7 0.77 Substrate model best 1.6 90 9 0.69

Table 7.3. The metrics of the P-gp database screening hit list.

185

Model Enrichment (E) Yield (Y%) Coverage (A%) GH-Score Inhibitor model 1 2.46 0.4 25 0.06 Inhibitor model 2 2.44 0.4 30 0.07 Substrate model 40.7 7 7 0.07

Table 7.4. The metrics of the merged Maybridge database screening hit list.

186

Substrate Predicted IC50 Predicted IC50 Substrate model Experimentally (Inhibition model (Inhibition model Molecule (this study) confirmed (225) 1) 2) NSC156625 Y Y 7 5.4 NSC259968 Y Y 4.4 64 NSC328426 Y Y 1.8 2.5 NSC353076 N N/D 4.7 0.66 NSC354975 N N/D 7.5 13 NSC359449 N Y 4.7 0.66 NSC363997 N Y 7.9 620 NSC374980 N N/D 8.3 8.3 NSC618757 N Y 8.3 54 NSC630678 Y N/D 4.4 27 NSC634791 Y N/D 3.7 25 NSC636679 N N/D 48 Did not fit NSC646946 N Y 6.2 22 NSC651727 Y N/D 3.8 3.2 NSC676864 N N/D 8.8 2.9 NSC682066 N N/D 10 96 NSC694268 N Y 8.9 21 N/D = not determined

Table 7.5. The predictions for NSC compounds.

187

2 0

0 9 1 6 3 6 1 H 1 2 1

D B

4 2 1 0 1 H 2 2 1 2

B 6 0

4 1 1 7 6 6 6 1 6 1 R

acophores as well som odel

t t t i fi fi f e m

t at o r

not not n

7 5 1 3 5 6 d d d

8 7 8 6 3 0 i i i t ...... i 3 0 3 3 D D 1 3 Subst f D 2 50 odel IC r m o t icted bi

ed 7 5 2 5 5 3 8 7 . . . . . 0 . . r . 3 4 1 5 2 2 1 1 Inhi p 2 . Scores are shown for all pharm

t r odel 2 o 2 Fi t

bi 1 3 2 4 8 9 8 3 6 5 4 0 odel 3 6 7 9 8 6 ...... 6 6 7 6 6 5 6 6 6 Inhi m 1 50 odel IC r m o t icted bi

ed 9 4 9 8 5 2 7 9 1 ...... r 6 7 3 8 7 6 6 8 7 Inhi p

t r ber (RBN), H-bond donors (HBD) and acceptors (HBA). o 1 Fi t

bi 2 9 8 2 9 4 7 2 1 6 5 8 5 odel 5 6 6 5 6 ...... 2 2 2 2 2 2 2 2 2 Inhi m SCUT database search with inhibitor m

L O

N L FER

L E E I A O L L O

T T I O O R N N S R I N I R Z E ALC I O A L T E T A L S L E T E I I N G

L e M O M A O ITR L P L SOPR C T I I I E HOLEC E A Nam AC B C M M NAFC R S T Table 7.6. Hit list from descriptors for rotable bond num 188 HBD HBA MW RBN a. 4.2 13.7 581.6 12.0 b. 2.1 14.7 553.0 8.3 c. 2.7 16.1 655.7 10.9 d. 10.0 37.0 929.5 8.3 e. 5.0 35.0 1201.8 15.0

Table 7.7. Average properties of each section (a – e) in Figure 7.3b

189

Figure 7.1. Database scoring metrics (redrawn from (223)). a. D: the total number of compounds in the database; A: the number of active compounds in the database Ht: the number of compounds in a search hit list; Ha: the number of active compounds in the hit list b. Enrichment (E): indicates the ratio of yield of active compounds in the hit list relative to the yield of active compounds in the database. c. Yield (%Y): is the percentage of known active compounds in the hit list. d. Coverage (%A): is the percentage of known active compounds retrieved from the database. e. The Güner-Henry score is a combination of Yield and Coverage with the correction of hit list size. To be more accurate, the correction is the ration of retrieved false positive hits over inactive compounds in a database. Should the whole database be retrieved as a hit list, this correction will add a penalty since this won't be reflected by either Yield or Coverage. Small but selective hit lists (high %Y but low %A) are very useful because fewer false positive hits in the hit list can be considered an important lead. So Yield is more important than Coverage and was given a weight of 0.75 and Coverage has 0.25 weight in the above formula.

190

Figure 7.1.

191

Figure 7.2 Alignment of test molecules to the P-gp pharmacophore model. a. Pharmacophore model for inhibitor model 2 aligned with CP114416. Green indicates HBA feature and cyan indicates HYD feature. b. Molecules selected with the P-gp inhibitor model 2 for in vitro testing aligned with inhibitor model 2.

192

Figure 7.3. Venn diagrams showing the overlap of known P-gp substrates selected from the P-gp database (a) and the SCUT database (b) following searches using the three pharmacophore models.

193

CHAPTER 8

A LIGAND-BASED APPROACH TO IDENTIFY QUANTITATIVE

STRUCTURE-ACTIVITY RELATIONSHIPS FOR THE ANDROGEN RECEPTOR

8.1 Introduction

Selective Androgen Receptor Modulators (SARMs) were discovered to produce desirable anabolic responses while minimizing androgenic effects (230, 231). Recent in vivo studies in rats demonstrated that two novel SARMs (denoted S-1 and S-4) elicit potent and tissue selective pharmacologic effects. These nonsteroidal SARMs have many theoretical advantages over conventional steroid therapies, which are not selective between anabolic and androgenic tissue, have poor oral bioavailability, and are associated with side effects from cross-reactivity with other steroid receptors (232). Since the majority of nonsteroidal androgen receptor (AR) ligands demonstrate binding affinities

10-fold lower than the leading androgenic steroids (233), efforts were sought to improve upon the favorable receptor-ligand interactions to obtain more potent compounds. The

194

traditional medicinal chemistry and pharmacologic approach to design, synthesize, and

evaluate a large number of derivatives with varying AR binding affinities have resulted in

the refinement of structure-activity relationships (SAR) for AR binding. However, this

approach is both time- and cost-intensive and is not amenable to high throughput

screening. An integrated approach using molecular modeling (SYBYL 6.8) and CoMFA

was employed to avoid these constraints, to create a visual database incorporating our

ligands and other structurally diverse compounds, and to provide rationale for the

synthesis of higher affinity AR ligands with improved activity.

Structure activity relationships have been previously described for analogs of

hydroxyflutamide and bicalutamide (234). These include an electron-withdrawing group

at the 4-position of the A-ring and an amide linkage attached to the chiral center. The

chiral center favors the R configurations in S-linked and SO2-linked bicalutamide analogs

(235) and the S configuration in N and O-linked analogs (233) due to the lower priority of the nitrogen and oxygen with respect to the amide. Thus, all high affinity N, O, S, and

SO2-linked bicalutamide derivatives are of the same three-dimensional (3D)

configuration. The presence of a NO2 at the para-position of the A-ring appears to be

favored over a CN group due its stronger electron-withdrawing properties (236).

Waller et al. (237) using steroidal ligands, hydroxyflutamide, and a number of

pesticides developed a CoMFA model with a cross-validated r2 of 0.792 from a training

set of 20 compounds. Alignment was performed via overlap of the A-ring of DHT to 6-

membered rings found in each of the other ligands with an additional energy

minimization after field fit and internal coordinate adjustments to provide maximal

overlap with DHT (237). Steric bulk was favorable at several sites of the steroidal ring

195

system, including the B-ring at the C6 and C7 positions, C-ring at the C11 and C12 position, and D ring at the C17 position of steroids. In addition, a positive charge at the

C17 positions and negative charge at the C3 position of the steroid ring system resulted in increased affinity (237). Steric bulk at positions more distant from the C17 position was found to be unfavorable. Our approach to design a CoMFA model differed considerably for two reasons: (1) the receptor was used as a template to align diverse structures and (2) the majority of the structures used in our model were bicalutamide and hydroxyflutamide analogs. These modifications allow us to combine docking information into our ligand predictions as well as provide a model that is specific for ligands resembling hydroxyflutamide and bicalutamide.

The final CoMFA model is dependent on the 3D conformation of the individual molecules in the training set, as well as their overlap. Since the AR ligands synthesized thus far are conformationally mobile, the question becomes: how do we assign significance to one potential conformation relative to another? Originally, 3D quantitative SAR (QSAR) techniques such as CoMFA were developed for ligand design problems in the absence of structural information about the target receptor. However, it is increasingly applied in situations in which information regarding receptor structure exists in the form of biophysical data (X-Ray, NMR, etc.) or theoretical models (238). In these situations, a firm basis for conformational selection exists and can be expected to improve the overall effectiveness of the CoMFA-based ligand design approach. While the crystal structure for the AR ligand binding domain (LBD) is available, most studies suggest that bicalutamide and its analogs adopt a unique conformation in solution. Poujol et al.(239) suggested that the aromatic ring of bicalutamide interfered with appropriate

196

AR LBD folding. We developed a homology model of the AR to the progesterone receptor that allows docking of these compounds and explains the mechanism of binding for chiral nonsteroidal ligands (240). Importantly, this model also provides a scaffold for alignment purposes. Studies herein utilize this approach to superimpose the CoMFA contours and the receptor LBD to better understand the amino acid interactions that are important to obtain high AR binding affinity. However, it is important to note that the activity of the SARMs that we recently reported cannot be explained by the crystal structure, suggesting that it may be inaccurate due to crystal packing effects. Thus, the current work is also aimed at providing a rationale for site-directed mutagenesis in order to further investigate the amino acid-contour relationship and the structure of the AR

LBD during interaction with SARMs.

8.2 Materials and Methods

Competitive Binding Assay

Binding data for bicalutamide and its derivatives were determined in Dr. Dalton’s laboratory by a method previously described (233). Cytosolic AR was obtained from ventral prostates of Sprague-Dawley rats. Prostates were excised and immediately immersed in ice-cold homogenization buffer (10 mM Tris, 1.5 mM disodium EDTA, 0.25

M sucrose, 10 mM sodium molybdate, and 1 mM PMSF). Prostates were minced with scissors, homogenized, and centrifuged (Model L8-M, Beckman Instruments Inc., Palo

Alto, CA). The supernatant (cytosol) was collected and stored at – 80ºC until use.

Aliquots (50 µL) of AR cytosol were incubated with a saturating concentration of 3H-

197

MIB (1 nM) and 1 µM of triamicinolone acetonide at 4ºC for 18 hours in the absence or

presence of increasing concentrations of the compound of interest (10–1 nM to 104 nM).

Nonspecific binding of 3H-MIB was determined by adding 1000 nM MIB to the incubate.

Bound and free radioligand were separated using hydroxyapatite and the concentration of

3 bound H-MIB was determined. The IC50 (concentration of the test compound that

inhibited the specific binding of 3H-MIB by 50%) was determined using nonlinear regression. Relative binding affinity (RBA) was calculated as RBA = IC50 of DHT/ IC50 of compound of interest. Binding data for the steroidal ligands and tricyclic quinolinones were obtained from publications by Waller et al. (237) and Zhi et al. (241), respectively.

Only data with a reported DHT value was used in this study to normalize for inter-

laboratory variations. Biological data for the CoMFA model was input as pRBA.

Molecular Structure Building

The correct conformation of the ring structures was used during the structure

building process. AM1 Hamiltonian within the MOPAC suite of programs was used to

assign the point charges of atoms as well as to optimize the geometry by minimizing the

energy. Molecular mechanics corrections were made to any CO-NH bond using the

keyword MMOK. The keyword PRECISE, which increases the criterion for terminating

all optimizations by a factor of 100, was also used. Of 122 ligands, 10 were randomly

selected as test compounds; the remaining 112 ligands were used as the training data set.

198

Ligand Alignment

An important consideration in the creation of a CoMFA model is how to align the different compounds. Often, similar functional groups or common atoms are chosen as alignment points to determine how compounds will overlap. Docking ligands to a receptor has also been employed as a rationale for alignment. In the present study we used a combination of these methods by first solving the theoretical AR-bound conformations of several template AR ligands and then fitting the analogous molecules in the CoMFA training set to these template conformations using common atoms as alignment points. Compound S-4 (compound 5), DHT, and compound 85 were used as templates for the hydroxyflutamide derivatives, steroids, and tricyclic quinolinones, respectively. The templates for the two-ring model were the ether-linked compound S-4, the sulfone-linked bicalutamide, a secondary amine-linked derivative, and a thioether- linked derivative (see Figure 8.1 for alignment points). These representative ligands were docked into the AR LBD homology model (detail below) to determine their relative conformations. Analogous molecules were then fitted to their respective template ligands using SYBYL/MULTIFIT (242).

Two-ring ligands, such as bicalutamide, adopted a conformation very different from that of the MOPAC-optimized analogs. To avoid severely distorted structures, bicalutamide and bicalutamide-like compounds were rebuilt based on the conformation of the docked templates. A moderate energy minimization was performed (RMS gradient

1.0 and maximum steps 1,000) after structure generation in order to maintain the bound conformation without the receptor environment constraints. Alignment points were selected based on important functional groups and common atoms for each set of

199

molecules as shown in Figure 8.1. The final overlap was generated by merging all 6 sets of Fieldfit-aligned molecules and maintaining their relative docked position.

FlexX Docking

A model of the hAR LBD based on homology to the human progesterone receptor and postulated the binding modes for testosterone and several chiral nilutamide derivatives has been developed(240). In the current study, Michael Mohler from The

University of Tennessee docked a variety of molecules into the AR G1 homology model

(direct binding to ligand as opposed to water-mediated binding) using similar methodology as previously employed (240). Briefly, the docking was performed with

FlexX (243). The AR G1 model was prepared for docking by removing all hydrogens.

The ligand binding site was defined as all residues within 6.5 Å of testosterone. The dihedral angle orienting the hydroxyl of T877 was rotated to +120 degrees in the active site such that it could act as a H-bond donor with bound ligands. Ligands were docked into the ligand binding site with hydrogens present and formal charges assigned by FlexX.

The FlexX docking solutions were initially relaxed in FlexX and scores were assigned for both relaxed and unrelaxed docking solutions. The top 30 scoring (unrelaxed) docking solutions were saved.

Ligand-AR Complex Energy Refinement

Ligands that had multiple docking solutions were evaluated to determine those that would be chosen for further energy refinement. Docking solutions were prioritized by the C-Score of the relaxed conformations to eliminate all complexes except those for

200

which all four scoring functions predicted to be in the top 50% of the 30 docking solutions. Manual inspection of the ligand-AR interactions of these complexes allowed further elimination of complexes with fewer interactions. If multiple conformations remained, the final selection criterion was the best total score for the relaxed conformer.

Energy refinement was achieved by merging the docked ligand with the homology model

(without testosterone) and energy minimizing a subset of this ligand-receptor complex.

T877 was rotated to 120 degrees as described above, protons were added to the receptor, and MMFF94s charges were applied to the receptor for energy minimization. SYBYL

6.8 was employed to minimize the subset of this complex allowing the assignment of a hot radius (unrestrained movement) and an interesting radius (limited movement). The hot radius was 4.0 Å and the interesting radius was 8.0 Å from the ligand. The minimizations proceeded for 100,000 iterations or to a 0.005 kcal/mol termination gradient. The resulting ligand conformations served as the conformational templates (see above) for the CoMFA training sets.

CoMFA

The method is described in detail in the introduction chapter. pRBA was used as the biological activity is CoMFA study. Corticosterone was identified as the only outlier based on factor analysis. It was therefore excluded from the test set and the CoMFA model was regenerated.

201

8.3 Results

QSAR Validation

We used receptor binding and structural information from 122 ligands (Table 8.1) to identify important structure activity relationships for the AR interaction. RBA to DHT ranged from 0.02% to 20% for hydroxyflutamide analogs, 0.02% to 16% for bicalutamide analogs, 1.8% to 29% for tricyclic quinolinones, and 0.002% to 501% for steroids. Molecular modeling and CoMFA were used in a ligand-based approach in attempt to integrate unique aspects of ligand and receptor conformation and to better understand current and emerging data regarding AR structure and function. This approach resulted in a QSAR model that was highly predictive of RBA for the AR (Table

8.2). Conventional PLS analysis (Figure 8.2 a, b and Table 8.3) showed that actual and predicted pRBA were highly correlated (r2 = 0.974) with residual values randomly scattered around zero. The q2 was 0.593, corroborating the statistical validity of the

QSAR model. PRESS was calculated as 0.737, which is the predictive error sum of squares representing error in the cross-validated PLS correlation. The s value of 0.262 represents the sum of the squares error for the conventional PLS correlation. The CoMFA model comprises 39% steric and 61% electrostatic interactions. Corticosterone was excluded from the model because of its high residual suggesting that either the 11- or 21- hydroxyl group distinguishes it in terms of AR binding.

We used this model to predict the RBA of ten randomly selected compounds. The

RBA for these compounds ranged from 0.2% to 11% indicating that they represented a reasonable range of RBA observed in our studies. Further, each of the structural classes

202

of molecules was represented in this test set. A plot of actual vs. predicted pRBA for these compounds demonstrated a correlation of 0.953 (Figure 8.2c), corroborative of the predictive capability of the QSAR model. Residual pRBA values were randomly scattered around zero.

Steric and Electrostatic Interactions with the Ligand

Coefficient contour maps were used to examine the steric and electrostatic fields associated with AR binding. We focused our analysis on compound S-4. Previous studies in our laboratory (231) showed that this ligand is a potent and tissue-selective nonsteroidal androgen. Figure 8.3 illustrates the contours at the 5% and 30% contour level. Polyhedra in each map surround all lattice points and indicate points in the ligand that are most strongly associated with observed differences in RBA. The contours of the steric map (van der Waals forces) are shown in yellow and green, while the contours of the electrostatic map are shown in red and blue. Greater binding is correlated with less bulk near yellow, more bulk near green, more negative charge near red, and more positive charge near blue.

Contour levels were adjusted to 5% to describe the most important regions for high binding affinity (Figure 8.3a) and 30% to emphasize the contours present around the

B-ring (Figure 8.3b). At the 5% contour level, a red (negative charge) polyhedra near the nitro group of compound S-4 suggests that a H-bond acceptor is favored at the para- position of the A-ring. Substitution of the nitro group to cyano group maintains high binding affinity as would be expected. A green contour representing favorable hydrophobic interaction is seen at the 3-position of the A-ring. Compound S-4 contains a

203

trifluoromethyl group at this position, however, substitutions to iodine and chlorine maintained high binding affinity. At the ketone in the bridge of compound S-4, a blue contour is seen likely due to its conservation among all hydroxyflutamide and bicalutamide derivatives. A blue contour (favorable positive charge) at the hydroxyl group of the chiral carbon of compound S-4 indicates its role as an H-bond donor. Lastly, distant to the oxygen linkage of compound S-4 is a yellow contour representing steric hindrance at this region.

The 30% contour level exaggerates the polyhedra surrounding the B-ring of compound S-4. Since most molecules used in this model do not occupy this space, these contours are not present at high confidence. It is important to note that CoMFA using only bicalutamide derivatives (data not shown) could be used to display these interactions at lower contour levels. Both sides of the B-ring of compound S-4 are surrounded by yellow contours demonstrating the steric hindrance limiting the size of the substituent on the B-ring. The moderately sized acetamido group at the para-position can be accommodated according to the model. Bulky para substituents and moderately sized meta- and ortho- substituents, however, would result in unfavorable steric interaction according to the model.

Ligand and AR Interaction

We used an overlap of the AR homology model and the CoMFA model at the 5% contour level to identify the most important residue interactions for high binding affinity

(Figure 8.4a). The nitro group of compound S-4 at the para-position of the A-ring appears to interact with ARG752 and GLN711, taking advantage of this region of favorable

204

negative charge identified in the CoMFA model. The trifluoromethyl group at the meta-

position of the A-ring of compound S-4 is closely surrounded by VAL746 and MET742 suggesting that these residues are responsible for favorable hydrophobic interactions with the ligand. The ASN705 overlaps with a blue contour suggesting that the Oδ1 of ASN705 accepts the hydrogen from chiral hydroxyl group of compound S-4. Only isomers allowing this interaction have high affinity for AR. Steric hindrance seen distant to the oxygen linkage group of S-4 overlaps with THR877.

The CoMFA model at a 30% contour level overlapped with the homology model

and aids in identification of amino acids bordering the B-ring (Figure 8.4b). According to

the homology model, the B-ring lies in a subpocket bordered by MET780, CYS784, and

MET787. A predominance of steric hindrance at this position is well supported by MET780 and MET787, which appear to border the acetamido group at the para-position of

compound S-4. These methionine residues also portray why moderately sized meta-

substituents result in poor binding affinity. Bulky residues at the para-position also

decrease binding affinity likely due to unfavorable steric interaction with the residues of

this subpocket.

8.4 Discussion

The model described herein significantly advances our understanding of AR-

ligand interactions. We incorporated a large number of chemically diverse ligands in an

integrated approach using CoMFA and homology modeling. A previous CoMFA study

by Waller et al. (237) used the A-ring of steroids and hydroxyflutamide for alignment.

205

However, our homology modeling and docking studies of the AR suggest that the A-ring

of steroids lies in a different region (Figure 8.5). The 3-keto group of DHT and the NO2 group of bicalutamide-like derivatives share a similar space. At the 2-position, the trifluoromethyl group of S-4 and the ring system of steroids overlaps with the methyl group of tricyclic quinolinones in the region shown to favor hydrophobicity. The 17-OH of DHT as shown overlaps close to the linkage region of S-4; however, the hydroxyl group of S-4 likely substitutes for its role as an H-bond donor. As visualized in this overlap, S-4 occupies an additional space in the receptor not identifiable with steroids, hydroxyflutamide analogs, and tricyclic quinolinones.

Favorable negative charge on the ligand at the 4-position of the A-ring of bicalutamide and hydroxyflutamide analogs is supported by the location of ARG752 and

GLN711 in the homology and crystal structures of the AR, and the activity of the receptor

mutants. These residues act as H-bond donors in the crystal structure to the 3-keto group

of R1881 (244) and in the homology model to the 4-nitro group of hydroxyflutamide

(240). Substitution of a cyano group at this position in bicalutamide derivatives maintains

high AR binding affinity. Mutations of ARG752 (e.g. R752Q mutants observed in partial

androgen insensitivity syndrome (PAIS)) result in loss of this interaction and poor

binding affinity (244).

The blue contour above the chiral hydroxyl group of bicalutamide analogs

indicates the interaction with ASN705. This amino acid is also important for ligand

discrimination and binding. N705S mutants observed in complete androgen insensitivity

syndrome (CAIS) corroborates the importance of this interaction (244). This area

overlaps with the 17-position of steroids (240). Poujol et al. (239) demonstrated near

206

complete loss of the antagonist properties of hydroxyflutamide and bicalutamide in the

N705A mutation, while RU486 and cyproterone acetate maintained the ability to repress transcriptional activation (239). This suggests that the interaction with hydroxyflutamide and bicalutamide derivatives is essential for binding. Further, the published crystal structure of the AR demonstrates that both ASN705 Oδ1 and THR877 Oγ are H-bonded to

R1881 [16]. The chiral hydroxyl group of bicalutamide and hydroxyflutamide does not interact with THR877 according to their docked conformation.

The yellow contour flanking the oxygen linkage group of S-4 overlaps with

THR877. Similar contours were seen in the model by Waller et al (237). This is an

important discriminatory region in AR disfavoring interactions with steroids (e.g.

mineralocorticoids, estrogens, progestins) that contain bulkier groups at the 17-position.

The T877A mutation as is observed in the LNCaP cancer cell line allows ligands with

more bulky constituents at this position to bind the AR and results in less specificity in

binding (245).

A hydrophobic favorable region seen in green at the 2-position of the A-ring in

hydroxyflutamide and bicalutamide analogs is present with high confidence as well. This

position corresponds to carbons 5, 6, and 7 in steroids and a methyl group in tricyclic

quinolinones. In hydroxyflutamide and bicalutamide analogs, substitutions of chlorine,

and iodine for the trifluoromethyl group at this position sustained high binding affinity, while the substitution of a hydrogen resulted in poor binding affinity to the AR. VAL746

and MET742 appear to be responsible for this contour, as can be easily visualized in the overlap of the receptor and the CoMFA model. Mutations of these residues have been noted in PAIS further supporting their importance (232, 246).

207

The region bordering the B-ring of the bicalutamide analogs is an important aspect of the CoMFA model due to the variation between the homology model and the crystal structure in this region (244). Since only bicalutamide derivatives occupy the space of the bicalutamide B-ring, contours were set to 30% to exaggerate the contours in this area. Steric contours closely border the B-ring. These interactions appear to be mediated chiefly by MET787 and MET780 upon comparison to the homology model.

Similar contours are seen at higher confidence levels when performing CoMFA with a data set incorporating only bicalutamide-like derivatives. Alignment to the crystal structure was not possible due to the tighter packing amino acids in this pocket in the crystal structure as compared to the homology model. MET780 in the crystal structure appears to forbid the docked conformation of bicalutamide analogs (Figure 8.6).

Functional groups such as the acetamido group can only be accommodated at the para- position of on the B-ring. The current study indicates that this is due to steric hindrance from hydrophobic residues surrounding this area of the binding pocket.

It appears from the model that the highest affinity SARMs synthesized in our laboratory share all of the common electrostatic interactions with the highest binding steroidal androgens. Androgenic steroids such as DHT have more than ten times the affinity for the AR compared to the SARMs developed in our laboratory. This appears to be most likely from additional hydrophobic interactions with the AR binding pocket.

Manipulating the hydrophobicity of certain regions of our SARMs to exploit some of these interactions might be key in increasing binding affinity. CoMFA will continue to play a role as such compounds are synthesized and tested.

208

8.5 Conclusion

The presented CoMFA model for the AR portrays a high predictive capacity for

the binding of AR ligands. Further, interactions identified using the CoMFA contours

correspond appropriately to important amino acid contact sites identified by molecular

modeling, the crystal structure, and mutational analysis. The use of bicalutamide analogs,

which occupy a larger and apparently novel space in the LBD as compared to other

ligands (e.g. steroids, tricyclic quinolinones, and hydroxyflutamide analogs), provides

insight to a previously unknown region in the binding pocket of the AR. Steric contours

closely surrounding each side of the B-ring suggest that both MET787 and MET780 play an important role in preventing bulky B-ring substituents from binding the AR. This is the first study to suggest that these amino acids may be important for AR ligand binding.

Ongoing studies in our laboratory are targeted at elucidating the importance of these residues for the binding of bicalutamide-like derivatives and SARMs. We also showed that the CF3 position of nonsteroidal AR ligands is a site that favors hydrophobic interaction. Lastly, other documented interactions, such as the H-bond acceptor at the 3- keto and H-bond donor at the 17-OH group of steroidal androgens, support the CoMFA model. In summary, we used an integrated approach with CoMFA and homology modeling to investigate important interactions for AR ligand binding.

209

R6 R R 5 R7 1 O * R2 N X R8 R H 3 R4 R9

cpds actual predicted residual R1 R2 R3 R4 X R5 R6 R7 R8 R9 isomer

1 1.35 1.36 -0.01 NO2 CF3 OH CH3 O H H F H H S

2 2.92 2.63 0.29 NO2 CF3 OH CH3 O H H F H H R

3 2.13 1.88 0.25 NO2 CF3 OH CH3 O H H COCH3 H H S

4 1.35 1.55 -0.2 NO2 CF3 OH CH3 O H H COCH2CH3 H H S

5 1.17 1.18 -0.01 NO2 CF3 OH CH3 O H H NHCOCH3 H H S

6 0.96 1.43 -0.47 CN CF3 OH CH3 O H H F H H S

210 7 0.98 1.41 -0.43 CN CF3 OH CH3 O H H Cl H H S

8 0.75 1.12 -0.37 NO2 CF3 OH CH3 O H H F F H S

9 1.51 0.95 0.56 NO2 CF3 OH CH3 O H H F F F S

10 1.47 1.19 0.28 NO2 CF3 OH CH3 O H F H F F S

11 1.24 1.18 0.06 NO2 CF3 OH CH3 O H H F H CH3 S

12 0.96 1.12 -0.16 NO2 CF3 OH CH3 O H H F H F S

13 1.23 0.89 0.34 NO2 CF3 OH CH3 O H H H H F S

Continued

Table 8.1. Structures and predictions of AR CoMFA training set

Table 8.1 continued

14 0.88 0.95 -0.07 NO2 CF3 OH CH3 O F F F F F S

15 1.04 1.17 -0.13 NO2 CF3 OH CH3 O H H Cl F H S

16 1.37 1.09 0.28 NO2 CF3 OH CH3 O H H F Cl H S

17 0.69 1.14 -0.45 NO2 CF3 OH CH3 O H H Cl Cl H S

18 1.67 1.33 0.34 NO2 CF3 OH CH3 O H H NHCOCH2Br H H S

19 0.8 1.03 -0.23 NO2 CF3 OH CH3 O H H NHCOCH2Cl H H S

20 1.61 1.83 -0.22 Cl CF3 OH CH3 O H H F H H S

21 1.96 1.7 0.26 F CF3 OH CH3 O H H F H H S

22 1.1 1.26 -0.16 CN Cl OH CH3 O H H F H H S

23 2.16 1.86 0.3 Br CF3 OH CH3 O H H Cl H H S

24 1.61 1.85 -0.24 CN CF3 OH CH3 SO2 H H F H H R

211 25 3.13 2.95 0.18 CN CF3 OH CH3 SO2 H H F H H S

26 2.53 2.4 0.13 CN CF3 OH CH3 S H H NH2 H H R

27 3.74 3.94 -0.2 CN CF3 OH CH3 S H H NH2 H H S

28 2.07 2.3 -0.23 CN CF3 OH CH3 SO2 H H NH2 H H R

29 1.26 1.5 -0.24 CN CF3 OH CH3 S H H NHCOCH3 H H R

30 3.71 3.09 0.62 CN CF3 OH CH3 S H H NHCOCH3 H H S

31 2.59 2.46 0.13 CN CF3 OH CH3 S H H N(COCH3)2 H H R

32 1.77 1.7 0.07 CN CF3 OH CH3 SO2 H H NHCOCH3 H H R

33 3.39 3.51 -0.12 CN CF3 OH CH3 S H H N(COCH2CH3)2 H H R

34 2.61 2 0.61 CN CF3 OH CH3 S H H NHCOCH2Br H H R

35 3.74 4.06 -0.32 CN CF3 OH CH3 S H H H NHCOCH2Br H R

Continued

Table 8.1 continued

36 0.8 1.17 -0.37 CN CF3 OH CH3 S H H NHCOCH2Cl H H R

37 3.18 2.95 0.23 CN CF3 OH CH3 S H H NHCOCH2Cl H H S

38 1.41 1.67 -0.26 CN CF3 OH CH3 SO2 H H NHCOCH2Cl H H R

39 2.2 2.17 0.03 CN CF3 OH CH3 SO2 H H H NHCOCH2Cl H R

40 2.95 2.53 0.42 CN CF3 OH CH3 S H H H NCS H R

41 3.55 3.97 -0.42 CN CF3 OH CH3 S H H H NCS H S

42 3.74 3.71 0.03 CN CF3 OCOCH3 CH3 S H H N(COCH2CH3)2 H H R

43 2.32 2.07 0.25 NO2 CF3 OH CH3 S H H NH2 H H R

44 3.29 3.54 -0.25 NO2 CF3 OH CH3 S H H NH2 H H S

45 1.03 1.07 -0.04 NO2 CF3 OH CH3 S H H NHCOCH3 H H R

46 2.66 2.7 -0.04 NO2 CF3 OH CH3 S H H NHCOCH3 H H S

212 47 1.54 1.25 0.29 NO2 CF3 OH CH3 SO2 H H NHCOCH3 H H R

48 0.98 1.14 -0.16 NO2 CF3 OH CH3 S H H NHCOCF3 H H R

49 1.36 1.1 0.26 NO2 CF3 OH CH3 SO2 H H NHCOCF3 H H R

50 1.09 0.84 0.25 NO2 CF3 OH CH3 S H H NHCOCH2Cl H H R

51 2.56 2.7 -0.14 NO2 CF3 OH CH3 S H H NHCOCH2Cl H H S

52 1.41 1.71 -0.3 NO2 CF3 OH CH3 SO2 H H NHCOCH2Cl H H R

53 2.94 3.06 -0.12 NO2 CF3 OH CH3 SO2 H H NHCOCH2Cl H H S

54 2.27 2.01 0.26 NO2 CF3 OH CH3 S H H NHSO2CH3 H H R

55 1.74 1.46 0.28 NO2 CF3 OH CH3 SO2 H H NHSO2CH3 H H R

56 1.55 1.52 0.03 NO2 CF3 OH CH3 O H H Cl H H S

57 2.05 1.93 0.12 NO2 CF3 OH CH3 O H H I H H S

Continued

Table 8.1 continued

58 1.63 1.76 -0.13 NO2 CF3 OH CH3 O H H Br H H S

59 2.11 1.78 0.33 NO2 CF3 OH CH3 O H H CH3 H H S

60 1.47 1.43 0.04 NO2 CF3 OH CH3 NH H H F H H S

61 0.99 1.19 -0.2 CN I OH CH3 O H H F H H S

62 1.85 2.22 -0.37 CN CF3 OH CF3 S H H N02 H H R

63 1.87 2.1 -0.23 NO2 CF3 OH CF3 S H H N02 H H R

64 1.27 1.17 0.1 NO2 CF3 OH CF3 S H H NHCOCH2Cl H H R

65 1.27 1.52 -0.25 NO2 CF3 OH CF3 SO2 H H NHCOCH2Cl H H R

66 1.38 1.38 0.0011 CN CF3 OH CH3 O H H NHCOCH3 H H R

67 3.09 3.24 -0.15 NO2 CF3 OH CH3 O H H NHCO2C(CH3)3 H H R

68 1.23 1.43 -0.2 NO2 CF3 OH CH3 O H H NCS H H R

213 69 2.68 2.52 0.16 NO2 CF3 OH CH3 NH H H NHCOCH3 H H R

Continued

Table 8.1 continued

R4 O * R3 N X R OH H 5 R2 R1

cpds actual predicted residual R1 R2 R3 R4 R5 X Isomer

70 1.47 1.54 -0.07 H CF3 CN H CH3 Br R

71 3.64 3.55 0.09 NO2 H NO2 H CH3 Br R

72 3.74 3.59 0.15 NO2 H NO2 H CH3 Br S

73 2.48 2.62 -0.14 H NO2 H NO2 CH3 Br R

74 3.74 3.57 0.17 H H CF3 H CH3 H N.A.

75 3.74 3.81 -0.07 CF3 H H H CH3 H N.A.

76 3.74 3.67 0.07 H H NO2 H CH3 H N.A.

77 3.74 3.63 0.11 NO2 H H H CH3 H N.A.

78 0.71 0.84 -0.13 H CF3 NO2 H CH3 Br R

79 1.12 1.39 -0.27 H CF3 NO2 H CH2CH3 CH3 N.A.

80 1.65 1.71 -0.06 H CF3 NO2 H CH3 CH3 R

81 1.58 1.78 -0.2 H CF3 NO2 H CH3 CH2CH3 R

82 2.05 2.15 -0.1 H CF3 NH2 H CH3 Br R

CF H 3 R5

R O NN2 R1 HH

cpds actual predicted residual R1 R2 R3 R4 R5

83 1.11 0.94 0.17 CH3 CH3 H H H

84 0.54 0.82 -0.27 CH3 H H H H

85 0.6 0.68 -0.08 CH2CH3 H H H H

86 1.74 1.45 0.29 CH(CH3)2 H H H H

87 0.48 0.66 -0.19 CH3 H H H CH3

88 1.06 0.76 0.3 CH3 H H H CH2CH3

89 0.65 1.05 -0.4 CH2CH3 H H H CH2CH3

90 1.02 1.14 -0.12 CH2CH3 H H H CH3

91 1.04 1.01 0.04 CH3 H H CH3 H

Continued

214

Table 8.1 continued

92 0.88 1 -0.12 CH3 H CH3 H H

93 1.61 1.1 0.52 CH3 H CH3 CH3 H

94 0.78 0.92 -0.15 CH2CH3 H H CH3 H

95 0.74 0.76 -0.01 CH2CH3 H CH3 H H

96 1.4 1.08 0.31 CH2CH3 H CH3 CH3 H

97 0.74 0.78 -0.04 CH3 H H CH2CH3 H

98 0.88 0.9 -0.02 CH3 H CH2CH3 H H

99 0.9 1.13 -0.23 CH2CH3 H H CH2CH3 H

100 1.16 1.01 0.15 CH2CH3 H CH2CH3 H H

101 1.32 1.53 -0.21 CH2CH3 H CH2CH3 CH2CH3 H

102 1.2 1.19 0.02 CH2CH3 H H CH2CH2CH3 H

cpds actual predicted residual

O ∆4-ANDROSTENEDIONE 1.7 1.52 0.18

COCH3 OH 17α-HYDROXYPROGESTERONE4.7 4.28 0.42

O OH DIHYDROTESTOSTERONE 0 -0.02 0.02

O H OH 5α-ANDROSTANE-3α,17β-DIOL 1.6 1.21 0.39

HO H OH ESTRODIOL 1.34 1.48 -0.14

CH3 MIBOLERONE -0.7 -0.77 0.07

O CH3

Continued

215

Table 8.1 continued

OH CH3 METHYLTRIENOLONE -0.7 -0.68 -0.02

COCH3 PREGNENOLONE 5 5.52 -0.52

HO COCH3 PROGESTERONE 4.7 4.32 0.38

O OH TESTOSTERONE 0.48 0.64 -0.16

216

actual predicted residual

NC NHCOCH3 O

F3C N O a H OH 1.55 1.34 0.21

NO2 O

CF3 N b H HO 1.64 1.88 -0.24

O2N OCH3 O

F C N O 3 H c OH 1.71 1.6 0.11

O2N NHCOCH3 O O F C N S 3 H d O 3.34 2.64 0.7

O2N NHCOCH3 O

F C N S 3 H e F3C OH 0.97 1.25 -0.28 SCN O H2 C F3C N Br H f OH 3.55 3.73 -0.18 H O2N O N

F3C N O g H OH 1.8 1.31 0.49

O2N NHCOCH2CH3 O

F3C N S H h OH 2 1.96 0.04

H O

H N I H OH 3.74 3.93 -0.19 CF3 H H H H

H O N N i-Pr j H H 1.08 0.62 0.46

Table 8.2. Test set structures and predictions

217

press q2 r2 s %Steric %ElectrostaticExclusion 0.737 0.593 0.974 0.262 39 61 Corticosterone

Table 8.3. Statistics for CoMFA model.

218

Figure 8.1. Alignment points for each set of compounds. (a) Conformation Alignment Points. Hydroxyflutamide analogs were aligned to the ether-linked SARM, S-4 (compound 5), by the points shown in green. Bicalutamide analogs were aligned using the pink and green points highlighted in S-4. Compounds 24, 29, S-4, and 60 were used as representatives of sulfonyl, thioether, ether, and amide-linked ligands, respectively for docking. DHT was docked into the homology receptor and other steroids were then aligned by the points shown in pink to DHT. Compound 85 was docked into the homology receptor and other tricyclic quinolinones were aligned to this ligand by the points shown in blue. (b) Overlap of ligands generated from field fit. Training and test compounds were aligned to their analogous structure by the points shown in panel (a) prior to CoMFA analyses.

219

Figure 8.2: Plot of residuals and predictions. (a) A plot of the actual vs. the predicted pRBA of the training set displays that compounds are closely scattered along the line with a slope of 0.949 and display a correlation of 0.974 as calculated by KaleidaGraph 3.5 (b) A plot of the residuals following exclusion of outliers for the training set. (c) A plot of the actual vs. the predicted pRBA for the test set demonstrates a correlation of 0.953. (d) Plot of the residuals from the test set.

220

Figure 8.3. CoMFA model and ligand interactions. (a) The contour plot shown at 5% contour levels demonstrates the most important areas of ligand interactions. (b) The 30% contour levels allow the contours surrounding the B-ring of compound S-4 to be visualized.

221

Figure 8.4. CoMFA contours at different levels for AR (a) CoMFA model displayed at 5% contour levels emphasizes critical residues for AR binding. (b) CoMFA model displayed at 30% contour levels emphasizes steric hindrance of the B-ring by MET780 and MET787.

222

Figure 8.5. Overlap of a S-4, DHT, and a tricyclic quinolinone from the docked conformation into the AR homology receptor. S-4 is colored by atom, DHT is colored as green, and tricyclic quinolinone is colored yellow. Compound S-4 requires an additional space in the receptor as compared with steroids and tricyclic quinolinones according to this model.

223

Figure 8.6. Overlap of AR crystal structure (colored by atom) and homology model (orange) binding pocket. The major difference between our homology model and the crystal structure is the position of MET780. Docking of bicalutamide analogs to the crystal structure will not occur due to this discrepancy.

224

CHAPTER 9

COMPARATIVE MODELING OF THE BILE ACID RECEPTOR

9.1 Introduction

The bile acid receptor (BAR) is formerly known as the Farnesoid X receptor

(FXR). It is a member of nuclear receptor family that transcriptionally regulates the production, movement and absorption of bile acid (247, 248). As the endogenous ligands for BAR, bile acids are the end products of hepatic cholesterol catabolism. Their production is intimately linked to cholesterol homeostasis. Due to its role in the regulation of bile acid and cholesterol levels, BAR is an attractive pharmaceutical target for treatment of hyperlipidemia diseases. BAR is expressed in the liver and intestine as well as other cholesterol-rich tissues. As a member of the nuclear receptor family, BAR contains a conserved central DNA-binding domain and a ligand-binding domain (LBD) that mediates ligand recognition, receptor dimerization and ligand-dependent activation

(249). The LBD functions as the molecular switch initiated by the binding of appropriate hormones. The knowledge of BAR LBD 3D structure will provide a molecular

225 explanation of how bile acids interact with LBD, thus opening the opportunity of rational drug design targeting lipid metabolism related diseases. For a more detailed review of ligand-based studies for BAR, please refer to (250). In this study, a BAR LBD comparative model was generated with satisfactory stereochemical properties. After the completion of this study, the BAR LBD crystal structure became available. Thus, a comparison of modeling result and experimental data was performed to provide useful insights into the application of de novo modeling studies.

9.2 Methods

Comparative modeling

The comparative modeling was performed using the program Modeller 7v7 (7).

BAR ligand binding domain (LBD) sequence was downloaded from Genebank. Modeller sequence_srch command was used to search best template among available PDB structures. In addition, Pfam database search (http://pfam.wustl.edu) was also performed.

All available template sequences were aligned according to their 3D structures using

MALIGN3D command. The resulting alignment, together with Pfam multiple sequence alignment, will provide guidance for later BAR LBD sequence alignment adjustment.

The sequences of available templates and BAR LBD were again aligned using Modeller

MALIGN, followed by ID_TABLE and DENDROGRAM command. The vitamin D receptor LBD (1ie8A, 28% sequence identity, 1.5 Å) and PPAR-gamma LBD (1fm9D,

26% sequence identity, 2.1 Å) were selected as final templates based on their sequence homology to BAR LBD as well as their high crystal structure resolution. The two

226 template structures were aligned first using MALIGN3D. The resulting alignment was aligned with BAR LBD sequence using ALIGN2D. The initial machine generated sequence alignment is shown in Figure 9.1. Multiple misalignments were reported by

CHECK_ALIGNMENT command. The sequence alignment was adjusted manually guided by the sequence alignment of nuclear receptor and the multiple structure alignment of BAR homologs. After all misalignments were corrected, comparative model was generated. Further manual adjustments were made to minimize the restraint violation warnings given out by the model generation process and new models were regenerated based on the improved sequence alignment. This process was repeated until there is no more improvement. The sequence alignment used to generate the final comparative model is shown in Figure 9.2. Optimizations were also performed using Modeller for resulting comparative models. The final model was selected among the twenty generated comparative models based on ProCheck evaluation results. A series of structure quality check programs were performed to validate the selected comparative structure.

Ramachandran plots generated by ProCheck for comparative model as well as the templates are shown in Figure 9.3. The percent of favored, allowed and disallowed Phi-

Psi angles are also listed. The model structure was then submitted to a more rigorous test program MatchMaker, which evaluates how well the sequence fits its 3D structure by assigning each amino acid an arbitrary energy. Lower total energies represent better fits

(251). As a control, fit energies were also calculated for template structures (Figure 9.4).

227

Flexible docking

Chenodeoxycholic acid (CDCA) is the cognate substrate of BAR and was docked into BAR model using flexible docking program FlexX (243). An array of cholic acids with different modifications were also attempted to dock into the model. The BAR active site was defined by overlapping BAR structure and the two template structures, and then define a pocket in the 6.5 angstrom radius around both template ligands. All hydrogen atoms were removed from the active site. Ligands were docked into the ligand binding site with hydrogen atoms present and formal charges assigned by FlexX. The docking results were optimized after being clustered. As a positive control for the docking protocol, two co-crystallized ligands were extracted from the templates and then docked into their active pockets. The overlaps of the docked and original ligand are shown in

Figure 9.5A-B.

9.3 Results and discussion

In this study, we derived a stereochemically correct comparative model for BAR

LBD. The Ramachandran plot (Figure 9.3) indicates that the comparative model has a stereochemical quality that is similar to that of the templates. The model has a higher energy (-0.07kT) than those of the templates (-0.21kT for 1fm9D and -0.14kT for 1ie8A), but this is a reasonable result to expect for a comparative model (252). Comparison of the

MatchMaker BAR energy graph and template energy graphs revealed higher energy at alpha I helix region (Figure 9.4) indicating a poor fit at this region.

FlexX successfully docked CDCA into the active site as shown in Figure 9.5C.

The rest of the cholic acids were also successfully docked into the cavity, except TCA,

228

DCA and TDCA. Among the failed dockings, TCA and TDCA have been previously

experimentally verified as poor substrates (52). As a further attempt to correlate the

docking results with in vitro data, the docking energies for each docked compounds were

tabulated with the experimental inhibition constant Ki values in Table 9.1. With five mismatches, no correlation was identified.

Upon completion of this study, the BAR LBD structure was crystallized and published by two separate groups (253, 254). This provided an excellent opportunity to validate the generated comparative model using experimental data. When fitted, the backbone RMSD between the model and crystal structure (1OSV) is 7 Å. The RMSD decreased to 3.1 Å after removal of alpha I helix (GLU241:LEU271), which is not aligned accurately. The poor fit between sequence and structure of alpha I helix had already been indicated by the MatchMaker (Figure 9.3). The structural overlap of the two structures is shown in Figure 9.6A-B.

The active sites share a large number of common residues. Not only do 12 out of

16 amino acids defined in crystal structure active site appear in our modeled BAR LBD active site (6 Å around docked CDCA), their relative positions also resemble each other.

(Figure 9.6C-D) The model is not without caveats though; the docked CDCA is 180 degrees reversed inside the binding pocket compared to the co-crystallized ligand. This is due to the unusual substrate orientation of BAR. Among all nuclear receptor family members with available structures, BAR is the only one that orients its substrate 180 degrees orthogonal to the ligands of the other nuclear receptor family members (253).

The comparative model was generated based on the information from all previous nuclear receptors and this restrained the substrate orientation to the available structures, missing

229 the actual orientation, which is an outlier when compared with the training set. Most computational methods train their parameters according to available crystal structures, which are again optimized by computational methods. This represents a bootstrapping approach, which gives a better score to the generated comparative models that are similar to those already available. This should always be taken into consideration when interpreting in silico models.

This study is a good example of a comparative model that is validated only through in silico methods, which even though these giving convincing results, was still significantly different to the biological structure determined by X-ray crystallography.

This could be because of the fact that most in silico methods are trained to fit available experimental data. Based on the same experimental data, in silico methods generally agree with each other. There is always hence the underlying danger of misinterpretation and false confidence when applying these in silico methods towards a system that is outside the training set. Caution should therefore be taken when interpreting in silico models, especially those not validated with in vitro methods.

230

Ligand FlexX docking energy (kJ/mol) Exp. Ki Mismatch? CDCA -8.374 (4.5uM) GCDCA -11.567 (10uM) TCDCA -10.324 (10uM) CA -4.992 (>1000uM) X GCA -12.511 (>1000uM) X TCA -11.75 (>1000uM) X DCA -9.95 (100uM) GDCA -11.674 (>500uM) X TDCA -9.937 (>500uM) X LCA -10.486 (3.8uM) GLCA -10.806 (4.7uM) TLCA -10.722 (3.8uM) cholesterol Failed non-ligand guggulsteroneE -11.737 inhibitor guggulsteroneZ -10.028 inhibitor

Table 9.1. Flexible docking results of different cholic acids (52).

231

1ie9A DSLRPKLSEEQQRIIAILLDAHHKTYDPTYSDF----CQFRPPVRVNDGGGSVTLELSQLSMLPHLAD 1fm9D PESADLRALAKHLYDSYIKSFPLTKAKARAILT----GKTTDKSPFVIYDMNSLMMGEDKIKFKHITP BAR EDSEGRDLRQVTSTTKSCREKTELTPDQQTLLHFIMDSYNKQRMPQEITNKILKEEFSAEENFLILTE

1ie9A LVSYSIQKVIGFAKMIPGFRDLTSEDQIVLLKSSAIEVIMLRSNESFTMDDMSWTCGNQDYKYRVSDV 1fm9D LQEQSKEVAIRIFQGCQFRSVEAVQEITEYAKSIPGFVNLDLNDQVTLLKYGVHEIIYTMLASLMNKD BAR MATNHVQVLVEFTKKLPGFQTLDHEDQIALLKGSAVEAMFLRSAEIFNKKLPSGH-----SDLLEERI

1ie9A TKAGHSLELIEPLIKFQVGLKKLNLHEEEHVLLMAICIVSPDRPGVQDAALIEAIQDRLSNTLQTYIR 1fm9D GVLISEGQGFMTREFLKSLRKPFGDFMEPKFEFAVKFNALELDDSDLAIFIAVIILSGDRPGLLNVKP BAR RNSGISDEYITPMFSFYKSIGELKMTQEEYALLTAIVILSPDRQYIKDREAVEKLQEPLLDVLQKLCK

1ie9A CRHPPPGSHLLYAKMIQKLADLRSLNEEHSKQYRCLSFQPECSMKLTPLVLEVFG------1fm9D IEDIQDNLLQALELQLKLNHPESSQLFAKLLQKMTDLRQIVTEHVQLLQVIKKTETDMSLHPLLQEIY BAR IH--QPENPQHFACLLGRLTELRTFNHHHAEMLMSWRVNDHKFTPLLCEIWDVQ------

1ie9A ---- 1fm9D KDLY BAR ----

Figure 9.1. Initial machine generated multiple sequence alignment among 1ie9A, 1fm9D and BAR

232

1ie8A --LRPKLSEEQQRIIAILLDAHHKTYDP------1fm9D PE----SA-DLRALAKHLYDSYIKSFPLTKAKARAILTGKTTDKSPFVIYDMNSLMMGE-DKIKFKHI BAR EDSEGRDLRQVTSTTKSCREKTELTPDQQTLLHFIMDSYNKQRMPQEITNKILKEEFSAE------

1ie8A ------TYSDFCQFRPPVRVNDGGGSVTLELSQLSMLPHLADLVSYSIQKVIGFAKMIPGFRD 1fm9D TPLQEQSKEVA------IRIFQGCQFRSVEAVQEITEYAKSIPGFVN BAR ------ENFLILTEMATNHVQVLVEFTKKLPGFQT

1ie8A LTSEDQIVLLKSSAIEVIMLRSNESFTMDDMSWT---CGNQDYKYRVS------DVTKAGHSLE 1fm9D LDLNDQVTLLKYGVHEIIYTMLASLMNK-D-GVLISEGQ--GFMTREFLKSLRKPFGD------BAR LDHEDQIALLKGSAVEAMFLRSAEIFNKKLPSGH------SDLLEERIRNS------GISDE

1ie8A LIEPLIKFQVGLKKLNLHEEEHVLLMAICIVSPDRPGVQDAALIEAIQDRLSNTLQTYIRCR-HPPP- 1fm9D FMEPKFEFAVKFNALELDDSDLAIFIAVIILSGDRPGLLNVKPIEDIQDNLLQALELQLK-LNHPESS BAR YITPMFSFYKSIGELKMTQEEYALLTAIVILSPDRQYIKDREAVEKLQEPLLDVLQKLCK-IHQPENP

1ie8A LLYAKMIQKLADLRSLNEEHSKQYRCLSF-QPECSMKLTPLVLEVFG--- 1fm9D QLFAKLLQKMTDLRQIVTEHVQLLQVIKKTETDMSL--HPLLQEIYKDLY BAR QHFACLLGRLTELRTFNHHHAEMLMSWR--VNDHKF--TPLLCEIWD-VQ

Figure 9.2. Final manually modified multiple sequence alignment among 1ie9A, 1fm9D and BAR

233

Core Allowed Generally allowed Disallowed BAR LBD 90.5% 8.6% 0.9% 0.0% VDR LBD 93.8% 5.8% 0.4% 0.0% PPAR-gamma LBD 88.4% 8.0% 2.0% 1.6%

Figure 9.3. Ramachandran plot of BAR LBD comparative model and the comparison of model and templates

234

Figure 9.4. MatchMaker energy plot for two templates, PPAR-gamma LBD (A), vitamin D3 receptor LBD (B) and BAR LBD model (C).

235

Figure 9.5. FlexX docking results. A. Overlap of FlexX docked ligand and crystallized ligand conformation for 1fm9D; B. The overlap of crystallized ligand and docked ligand inside VD3 receptor; C. Connolly surface view of the docked CDCA into BAR LBD model

236

Figure 9.6. Comparison of model result with crystal structure. A. Alignment of partial sequence comparative BAR LBD model and 1OSV (blue); B. Side view of the same alignment (90 degrees rotated); C. Ligand binding pocket from model structure; D. Ligand binding pocket from crystal structure.

237

CHAPTER 10

CONCLUSION

10.1 Summary and significance

Computationally generated pharmacophores and 3D-QSAR models have been widely applied in areas related to the discovery of novel molecules with affinity to a particular therapeutic target. However, the application to transporter problems, for which an X-ray crystal structure is not readily available, has been pioneered by our laboratory.

In this dissertation this approach has been applied to in vitro data for an array of transporters and receptors that can impact the absorption, distribution, and excretion of drug molecules. The generated models have not only provided insight into the binding features necessary for interaction with transporters and receptors, but have also been used to generate predictions for de novo molecules and screen databases for previously undiscovered substrates and inhibitors. The application of another in silico technique, comparative modeling, was also explored in this dissertation.

238

In the nucleoside transporter study (chapter 2), we examined the structural features of nucleoside transporter inhibitors and correlated these with their biological activity for the human homologs hCNT1, hCNT2 and hENT1 (33). 3D-QSAR models were calculated using two independent programs, i.e., CoMFA and GOLPE.

Pharmacophore models were also generated using DISCO. The thoroughly validated models provide guidance toward recognizing the molecular characteristics that are required for inhibiting each individual transporter isoform, their commonalities as well as specific features that set them apart. The models enabled us to predict transporter affinity and guide the design of novel lead compounds for drugs that may selectively target specific nucleoside transporter isoforms. Future studies would involve the analysis of a recently obtained dataset for hCNT3 that would complement the existing analyses.

When studying the peptide transporter (hPepT1, chapter 3), not only did we generated a Catalyst HIPHOP pharmacophore model, the model was also successfully applied in database screening to identify novel inhibitors with sub mM potency for hPepT1 (34). The pharmacophore was also able to identify known hPepT1 substrates and inhibitors in a database of over 500 commonly prescribed drugs. This study demonstrates the potential of combining computational and in vitro approaches to efficiently identify high potency inhibitors or substrates for important drug targets. When applied in a drug discovery setting, this novel approach could be utilized to rapidly identify transporter substrates or inhibitors prior to synthesis.

Both nucleoside transporter and hPepT1 play important roles in assisting drug absorption, which is an essential step before the drug can reach the target. The aforementioned 3D-QSAR and pharmacophore models for the two transporters provided

239 meaningful insight that could guide the design of drugs that are more easily transported into systemic circulation. This would enable lower drug doses for the same therapeutic effect due to improved oral bioavailability.

The first study that combined in vitro and in silico approaches to identify factors other than hydrophobicity that play a role in defining substrate interactions with human and rabbit OCT2, the 4th chapter described detailed structural features that contribute to the recognition of OCT2 (32). The CoMFA model also highlighted the multispecificity of the OCT2 binding site, and the permissiveness of the binding site with respect to steric bulk in other regions. Studies are currently under way to include other OCT orthologs and homologs.

With limited in vitro data for OATP, a novel method called meta-pharmacophore modeling, which incorporates biological data from different laboratories using various cell types, was applied and successfully generated statistically significant and predictive meta-pharmacophore models for rat Oatp1a1 and human OATP1B1 (36). The successful application of this novel approach expanded the scope of pharmacophore modeling to less well defined systems where available in vitro data is scarce. Instead of depending on an extensive set of in vitro results from one experimental setting, now pharmacophore modeling can be applied to newly identified systems with data obtained from various experimental sources.

Both OCTs and OATPs are involved in drug distribution and elimination and affect the bioavailability of a significant number of anionic and cationic drugs. Predictive models would assist us in estimating hepatic clearance and the renal secretory profile for

240 each given structure. Clearly, additional members of this important class of hepatic transporters need to be modeled to effectively predict hepatotoxicity.

The requisite structural features for P-gp mediated transport of a series of structurally related glucocorticoids were investigated in chapter 6. With the assistance of both pharmacophore and 3D-QSAR models, specific structural requirements of steroids for P-gp mediated transport were proposed. By applying the P-gp pharmacophore models as a screening tool in addition to other method and rules such as Lipinski’s Rule of Five in early drug discovery phase, expensive drug failures resulting from P-gp efflux related bioavailability problems could be reduced. This would assist the pharmaceutical industry with the “fail early, fail cheaply” paradigm.

Competition of multiple drugs at the same transporter can result in undesirable effects that are collectively known as drug-drug interactions. By identifying substrate binding requirements for each of these targets, possible drug-drug interactions at different levels and locations could be predicted early in drug discovery phase, thus avoiding costly failures due to unexpected drug-drug interactions in clinical phases (147, 148).

The application of in silico techniques are not limited to transporters, their application to the androgen receptor is exemplified in chapter 8. Integrated homology modeling and CoMFA studies allowed us to identify critical amino acids for ligand receptor interactions and provided QSAR data as the basis for mechanistic studies of androgen receptor structure, function, and design of optimized selective androgen receptor modulators.

Besides ligand based approaches, the comparative modeling technique was also explored in bile acid receptor study (chapter 9). Through careful sequence alignment,

241 even though not accurate, a stereochemically satisfactory model was developed. The comparison of in silico validation results and the crystal result which was later published emphasized the importance of continual validation using in vitro methods for in silico modeling results.

10.2 Evaluation of in silico approaches

Other than the biological significance carried by this dissertation study, the experience with multiple modeling programs also provides a unique opportunity for critical evaluation of each methodology.

Pharmacophores provide visual and qualitative information. They illustrate the three-dimensional nature of the ligand-receptor interaction by displaying important chemical features, a tolerance for each feature and distances between features. Their application in high throughput screening (HTS) is easily implemented since no structural alignment is needed. Their prediction is binary: either active or inactive. On the other hand, classic QSAR models and 2D–QSAR models are quantitative; however they lack any 3D information. Their implementation in HTS is also straightforward, thereby allowing activity prediction of each test compound. However, without conferring important 3D active conformation information of each ligand structure, the confidence in the predictions of these models is limited. As an improvement to classic QSAR models,

3D-QSAR models are quantitative and contain 3D information. However, due to their requirement of structural alignment of all compounds, it is difficult to apply 3D-QSAR model in HTS thus limiting its usefulness in drug discovery.

242

The Catalyst HypoGen pharmacophore is a successful approach in combining different aspects of ligand-based design. It both provides a quantitative prediction of compound activity and renders pharmacophore model in 3D. Being able to be applied in

HTS through either fast or best search algorithm makes the HypoGen pharmacophore approach a complete ligand-based drug design method. However, it is not without caveats.

The requirement of a wide activity range (4 – 5 orders of magnitude) limits the application of this approach.

10.3 Future perspectives

Clearly, the future of in silico drug discovery will lie in the realm of a combination of in silico approaches. The 3D-QSAR and pharmacophore techniques have proved their usefulness in different projects in this dissertation study. Multiple models may also be required for each protein to adequately predict affinity for different classes of molecules (108). These models may also be merged to show qualitative commonalities between different datasets. Data for the same protein generated in different in vitro systems can also be combined to result in what we have termed a Meta-pharmacophore approach (36). Further combining these ligand based approaches with receptor based methods such as comparative modeling could generate a multidimensional approach to derive a more complete picture of the binding process at molecular level (37). The integration of the models described above with other computational technologies such as de novo growth, docking algorithms and virtual library screening will assist in improving the rate of discovery of bioactive molecules (255) with optimal biopharmaceutical properties.

243

Ultimately these models could certainly be used to rapidly search databases to identify and remove undesirable molecules or suggest molecules that could be used as novel experimental probes for the protein in question. To date there have been few instances of such pharmacophore model-based-database searches (34, 221) and they represent a cherry picking approach to further filter compound vendor databases. The in vitro test results of predictions made by in silico models can be then incorporated back into the model for improved training set coverage. This integrated in silico and in vitro approach will ensure continued improvement of in silico models for more accurate prediction and more comprehensive prediction space coverage. Also, by actively combining the insights from in slico models with the interpretation of in vitro data, we will be able to gain a more complete understanding of the many transporters and receptors (256, 257).

Furthermore, the emergence of expression systems and high-throughput screening methods has allowed the collection of functional data on many transporter homologs and orthologs. The era of modeling individual transporters will now give way to comparative models of protein families that would illuminate the subtleties of genetic evolution to tweak substrate affinity of homologous proteins. With the more complete understanding of roles of different transporters and receptors in drug absorption, distribution, metabolism, excretion and toxicity (ADME/Tox) and enough in vitro data for each transporter and receptor involved, ultimately this dissertation study should assist in the prediction of ADME/Tox properties for each proposed structure based on a compilation of multiple models involving different proteins.

244

BIBLIOGRAPHY

(1) Grisshammer, R., and Tate, C. G. (1995) Overexpression of integral membrane proteins for structural studies. Q Rev Biophys 28, 315-422.

(2) Rosenbusch, J. P. (2001) Stability of membrane proteins: relevance for the selection of appropriate methods for high-resolution structure determinations. J Struct Biol 136, 144-57.

(3) Boobis, A., Gundert-Remy, U., Kremers, P., Macheras, P., and Pelkonen, O. (2002) In silico prediction of ADME and pharmacokinetics. Report of an expert meeting organised by COST B15. Eur J Pharm Sci 17, 183-93.

(4) Zhang, E. Y., Phelps, M. A., Banerjee, A., Khantwal, C. M., Chang, C., Helsper, F., and Swaan, P. W. (2004) Topology scanning and putative three-dimensional structure of the extracellular binding domains of the apical sodium-dependent bile acid transporter (SLC10A2). Biochemistry 43, 11380-92.

(5) Zhang, E. Y., Phelps, M. A., Cheng, C., Ekins, S., and Swaan, P. W. (2002) Modeling of active transport systems. Adv Drug Deliv Rev 54, 329-54.

(6) Leach, A. R. (2001) Molecular modelling, principles and applications, Prentice Hall.

(7) Sali, A., and Blundell, T. L. (1993) Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol 234, 779-815.

(8) Laskowski, R. A., MacArthur, M. W., Moss, D. S., and Thornton, J. M. (1993) PROCHECK: a program to check the stereochemical quality of protein structures. J. Appl. Cryst. 26, 283-291.

(9) Vriend, G. (1990) WHAT IF: a molecular modeling and drug design program. J Mol Graph 8, 52-6, 29.

245

(10) Abramson, J., Smirnova, I., Kasho, V., Verner, G., Kaback, H. R., and Iwata, S. (2003) Structure and mechanism of the lactose permease of Escherichia coli.[see comment]. Science 301, 610-5.

(11) Huang, Y., Lemieux, M. J., Song, J., Auer, M., and Wang, D. N. (2003) Structure and mechanism of the glycerol-3-phosphate transporter from Escherichia coli.[see comment]. Science 301, 616-20.

(12) Almqvist, J., Huang, Y., Hovmoller, S., and Wang, D. N. (2004) Homology Modeling of the Human Microsomal Glucose 6-Phosphate Transporter Explains the Mutations That Cause the Glycogen Storage Disease Type Ib. Biochemistry 43, 9289-97.

(13) Zuniga, F. A., Shi, G., Haller, J. F., Rubashkin, A., Flynn, D. R., Iserovich, P., and Fischbarg, J. (2001) A three-dimensional model of the human facilitative glucose transporter Glut1. J Biol Chem 276, 44970-5.

(14) Hruz, P. W., and Mueckler, M. M. (2000) Cysteine-scanning mutagenesis of transmembrane segment 11 of the GLUT1 facilitative glucose transporter. Biochemistry 39, 9367-72.

(15) Mueckler, M., and Makepeace, C. (2004) Analysis of transmembrane segment 8 of the GLUT1 glucose transporter by cysteine-scanning mutagenesis and substituted cysteine accessibility. J Biol Chem 279, 10494-9.

(16) Alisio, A., and Mueckler, M. (2004) Relative proximity and orientation of helices 4 and 8 of the GLUT1 glucose transporter. J Biol Chem 279, 26540-5.

(17) Mueckler, M., Roach, W., and Makepeace, C. (2004) Transmembrane segment 3 of the Glut1 glucose transporter is an outer helix. J Biol Chem 279, 46876-81.

(18) Seigneuret, M., and Garnier-Suillerot, A. (2003) A structural model for the open conformation of the mdr1 P-glycoprotein based on the MsbA crystal structure. J Biol Chem 278, 30115-24.

(19) Stenham, D. R., Campbell, J. D., Sansom, M. S., Higgins, C. F., Kerr, I. D., and Linton, K. J. (2003) An atomic detail model for the human ATP binding cassette transporter P-glycoprotein derived from disulfide cross-linking and homology modeling. FASEB Journal 17, 2287-9.

(20) Chang, G. (2003) Structure of MsbA from Vibrio cholera: a multidrug resistance ABC transporter homolog in a closed conformation. J Mol Biol 330, 419-30.

(21) Lee, J. Y., Urbatsch, I. L., Senior, A. E., and Wilkens, S. (2002) Projection structure of P-glycoprotein by electron microscopy. Evidence for a closed conformation of the nucleotide binding domains. J Biol Chem 277, 40125-31.

246

(22) Campbell, J. D., Koike, K., Moreau, C., Sansom, M. S., Deeley, R. G., and Cole, S. P. (2004) Molecular modeling correctly predicts the functional importance of Phe594 in transmembrane helix 11 of the multidrug resistance protein, MRP1 (ABCC1). J Biol Chem 279, 463-8.

(23) Ecker, G. F., Pleban, K., Kopp, S., Csaszar, E., Poelarends, G. J., Putman, M., Kaiser, D., Konings, W. N., and Chiba, P. (2004) A three-dimensional model for the substrate binding domain of the multidrug ATP binding cassette transporter LmrA. Mol Pharmacol 66, 1169-79.

(24) Campbell, J. D., Biggin, P. C., Baaden, M., and Sansom, M. S. (2003) Extending the structure of an ABC transporter to atomic resolution: modeling and simulation studies of MsbA. Biochemistry 42, 3666-73.

(25) Holm, L., and Sander, C. (1991) Database algorithm for generating protein backbone and side-chain co-ordinates from a C alpha trace application to model building and detection of co-ordinate errors. J Mol Biol 218, 183-94.

(26) McGuffin, L. J., Bryson, K., and Jones, D. T. (2000) The PSIPRED protein structure prediction server. Bioinformatics 16, 404-5.

(27) Rosenberg, M. F., Velarde, G., Ford, R. C., Martin, C., Berridge, G., Kerr, I. D., Callaghan, R., Schmidlin, A., Wooding, C., Linton, K. J., and Higgins, C. F. (2001) Repacking of the transmembrane domains of P-glycoprotein during the transport ATPase cycle. EMBO Journal 20, 5615-25.

(28) Jones, P. M., and George, A. M. (2002) Mechanism of ABC transporters: a molecular dynamics simulation of a well characterized nucleotide-binding subunit. Proc Natl Acad Sci U S A 99, 12639-44.

(29) Yates, C. R., Chang, C., Kearbey, J. D., Yasuda, K., Schuetz, E. G., Miller, D. D., Dalton, J. T., and Swaan, P. W. (2003) Structural determinants of P-glycoprotein- mediated transport of glucocorticoids. Pharm Res 20, 1794-803.

(30) Ekins, S., Kim, R. B., Leake, B. F., Dantzig, A. H., Schuetz, E. G., Lan, L. B., Yasuda, K., Shepard, R. L., Winter, M. A., Schuetz, J. D., Wikel, J. H., and Wrighton, S. A. (2002) Three-dimensional quantitative structure-activity relationships of inhibitors of P-glycoprotein. Mol Pharmacol 61, 964-73.

(31) Bednarczyk, D., Ekins, S., Wikel, J. H., and Wright, S. H. (2003) Influence of molecular structure on substrate binding to the human organic cation transporter, hOCT1. Mol Pharmacol 63, 489-98.

(32) Suhre, W. M., Ekins, S., Chang, C., Swaan, P. W., and Wright, S. H. (2005) Molecular Determinants of Substrate/Inhibitor Binding to the Human and Rabbit

247

Renal Organic Cation Transporters, hOCT2 and rbOCT2. Mol Pharmacol, 1067- 77.

(33) Chang, C., Swaan, P. W., Ngo, L. Y., Lum, P. Y., Patil, S. D., and Unadkat, J. D. (2004) Molecular requirements of the human nucleoside transporters hCNT1, hCNT2, and hENT1. Mol Pharmacol 65, 558-70.

(34) Ekins, S., Johnston, J. S., Bahadduri, P., D'Souza, V. M., Ray, A., Chang, C., and Swaan, P. W. (2005) In Vitro and Pharmacophore-Based Discovery of Novel hPEPT1 Inhibitors. Pharm Res 22, 512-7.

(35) Gebauer, S., Knutter, I., Hartrodt, B., Brandsch, M., Neubert, K., and Thondorf, I. (2003) Three-dimensional quantitative structure-activity relationship analyses of peptide substrates of the mammalian H+/peptide cotransporter PEPT1. J Med Chem 46, 5725-34.

(36) Chang, C., Pang, K. S., Swaan, P. W., and Ekins, S. (2005) Comparative Pharmacophore Modeling of Organic Anion Transporting Polypeptides:A Meta- analysis of Rat Oatp1a1 and OATP1B1. Molecular Pharmacology in press.

(37) Bohl, C. E., Chang, C., Mohler, M. L., Chen, J., Miller, D. D., Swaan, P. W., and Dalton, J. T. (2004) A ligand-based approach to identify quantitative structure- activity relationships for the androgen receptor. J Med Chem 47, 3765-76.

(38) Ekins, S., de Groot, M., and Jones, J. P. (2001) Pharmacophore and three dimensional quantitative structure activity relationship methods for modeling cytochrome P450 active sites. Drug Metab Dispos 29, 936-944.

(39) Cramer, R. D., Patterson, D. E., and D., B. J. (1988) Comparative Molecular Field Analysis (CoMFA). 1. Effect of Shape on Binding of Steroids to Carrier Proteins. J. Am. Chem. Soc. 110, 5959-67.

(40) Klebe, G., Abraham, U., and Mietzner, T. (1994) Molecular similarity indices in a comparative analysis (CoMSIA) of drug molecules to correlate and predict their biological activity. J Med Chem 37, 4130-46.

(41) Martin, Y. C., Bures, M. G., Danaher, E. A., DeLazzer, J., Lico, I., and Pavlik, P. A. (1993) A fast new approach to pharmacophore mapping and its application to dopaminergic and benzodiazepine agonists. J Comput Aided Mol Des 7, 83-102.

(42) Jones, G., Willett, P., and Glen, R. C. (1995) A genetic algorithm for flexible molecular overlay and pharmacophore elucidation. J Comput Aided Mol Des 9, 532-49.

248

(43) Clement, O. O., and Mehl, A. T. (2000) HipHop: Pharmacophore based on multiple common-feature alignments, in Pharmacophore perception, development, and use in drug design (Guner, O. F., Ed.) pp 69-84, IUL, San Diego.

(44) Li, H., Sutter, J., and Hoffman, R. (2000) HypoGen: An automated system for generating 3D predictive pharmacophore models, in Pharmacophore perception, development, and use in drug design (Guner, O. F., Ed.) pp 173-189, IUL, San Diego.

(45) van Drie, J. H. (2003) Pharmacophore discovery--lessons learned. Curr Pharm Des 9, 1649-64.

(46) Kristam, R., Gillet, V. J., Lewis, R. A., and Thorner, D. (2005) Comparison of conformational analysis techniques to generate pharmacophore hypotheses using catalyst. J Chem Inf Model 45, 461-76.

(47) Baroni, M., Costantino, G., Cruciani, G., Reganelli, D., Valigi, R., and Clementi, S. (1993) Generating Optimal Linear PLS Estimations (GOLPE): An Advanced Chemometric Tool for Handling 3D-QSAR Problems. Quant. Struct.-Act. Relat. 12, 9-20.

(48) Pastor, M., Cruciani, G., McLay, I., Pickett, S., and Clementi, S. (2000) GRid- INdependent descriptors (GRIND): a novel class of alignment- independent three- dimensional molecular descriptors. J Med Chem 43, 3233-43.

(49) Horwitz, J. P., Massova, I., Wiese, T. E., Besler, B. H., and Corbett, T. H. (1994) Comparative molecular field analysis of the antitumor activity of 9H- thioxanthen-9-one derivatives against pancreatic ductal carcinoma 03. J Med Chem 37, 781-6.

(50) Kroemer, R. T., Koutsilieri, E., Hecht, P., Liedl, K. R., Riederer, P., and Kornhuber, J. (1998) Quantitative analysis of the structural requirements for blockade of the N-methyl-D-aspartate receptor at the phencyclidine binding site. J Med Chem 41, 393-400.

(51) Schaal, W., Karlsson, A., Ahlsen, G., Lindberg, J., Andersson, H. O., Danielson, U. H., Classon, B., Unge, T., Samuelsson, B., Hulten, J., Hallberg, A., and Karlen, A. (2001) Synthesis and comparative molecular field analysis (CoMFA) of symmetric and nonsymmetric cyclic sulfamide HIV-1 protease inhibitors.[erratum appears in J Med Chem 2002 Jan 31;45(3):752]. Journal of Medicinal Chemistry 44, 155-69.

(52) Swaan, P. W., Szoka, F. C., Jr., and Oie, S. (1997) Molecular modeling of the intestinal bile acid carrier: a comparative molecular field analysis study. Journal of Computer-Aided Molecular Design 11, 581-8.

249

(53) Clark, M., and Cramer III, R. D. (1993) The Probability of Chance Correlation Using Partial Least Squares (PLS). Quant. Struct.-Act. Relat. 12, 137-145.

(54) Kroemer, R. T., and Hecht, P. (1995) Replacement of steric 6-12 potential- derived interaction energies by atom-based indicator variables in CoMFA leads to models of higher consistency. J Comput Aided Mol Des 9, 205-12.

(55) Bohacek, R. S., and McMartin, C. (1992) Definition and display of steric, hydrophobic, and hydrogen-bonding properties of ligand binding sites in proteins using Lee and Richards accessible surface: validation of a high-resolution graphical tool for drug design. J Med Chem 35, 1671-84.

(56) Kier, L. B., and Hall, L. H. (1986) Molecular connectivity in structure-activity analysis, Research Studies Press;

Wiley, Letchworth, Hertfordshire, England

New York.

(57) Pastor, M., Cruciani, G., and Clementi, S. (1997) Smart region definition: a new way to improve the predictive ability and interpretability of three-dimensional quantitative structure- activity relationships. J Med Chem 40, 1455-64.

(58) Kulkarni, S. S., Newman, A. H., and Houlihan, W. J. (2002) Three-dimensional quantitative structure-activity relationships of mazindol analogues at the dopamine transporter. J Med Chem 45, 4119-27.

(59) Kulkarni, S. S., Grundt, P., Kopajtic, T., Katz, J. L., and Newman, A. H. (2004) Structure-activity relationships at monoamine transporters for a series of N- substituted 3alpha-(bis[4-fluorophenyl]methoxy)tropanes: comparative molecular field analysis, synthesis, and pharmacological evaluation. J Med Chem 47, 3388- 98.

(60) Ravna, A. W., Sylte, I., and Dahl, S. G. (2003) Molecular model of the neural dopamine transporter. J Comput Aided Mol Des 17, 367-82.

(61) de Groot, M. J., Ackland, M. J., Horne, V. A., Alex, A. A., and Jones, B. C. (1999) Novel approach to predicting P450-mediated drug metabolism: Development of a combined protein and pharmacophore model for CYP2D6. J Med Chem 42, 1515- 1524.

(62) Baringhaus, K. H., Matter, H., Stengelin, S., and Kramer, W. (1999) Substrate specificity of the ileal and the hepatic Na(+)/bile acid cotransporters of the rabbit. II. A reliable 3D QSAR pharmacophore model for the ileal Na(+)/bile acid cotransporter. J Lipid Res 40, 2158-68.

250

(63) Jones, D. T., Tress, M., Bryson, K., and Hadley, C. (1999) Successful recognition of protein folds using threading methods biased by sequence similarity and predicted secondary structure. Proteins Suppl 3, 104-11.

(64) Xiang, Z., and Honig, B. (2001) Extending the accuracy limits of prediction for side-chain conformations. J Mol Biol 311, 421-30.

(65) Salas-Burgos, A., Iserovich, P., Zuniga, F., Vera, J. C., and Fischbarg, J. (2004) Predicting the three-dimensional structure of the human facilitative glucose transporter glut1 by a novel evolutionary homology strategy: insights on the molecular mechanism of substrate migration, and binding sites for glucose and inhibitory molecules. Biophys J 87, 2990-9.

(66) Gottschalk, K. E. (2004) Structure prediction of small transmembrane helix bundles. J Mol Graph Model 23, 99-110.

(67) Gottschalk, K. E., Soskine, M., Schuldiner, S., and Kessler, H. (2004) A structural model of EmrE, a multi-drug transporter from Escherichia coli. Biophys J 86, 3335-48.

(68) Ravna, A. W., Sylte, I., and Dahl, S. G. (2001) Molecular model of the Escherichia coli Na1/H1 antiporter NhaA. Receptors Channels 7, 319-28.

(69) Ravna, A. W., Sylte, I., and Dahl, S. G. (2003) Molecular mechanism of citalopram and cocaine interactions with neurotransmitter transporters. J Pharmacol Exp Ther 307, 34-41.

(70) Ekins, S., Kim, R. B., Leake, B. F., Dantzig, A. H., Schuetz, E., Lan, L. B., Yasuda, K., Shepard, R. L., Winter, M. A., Schuetz, J. D., Wikel, J. H., and Wrighton, S. A. (2002) Application of three dimensional quantitative structure- activity relationships of P-glycoprotein inhibitors and substrates. Mol Pharmacol 61, 974-981.

(71) Roman, D. L., Walline, C. C., Rodriguez, G. J., and Barker, E. L. (2003) Interactions of antidepressants with the serotonin transporter: a contemporary molecular analysis. Eur J Pharmacol 479, 53-63.

(72) Geldenhuys, W. J., Lockman, P. R., McAfee, J. H., Fitzpatrick, K. T., Van der Schyf, C. J., and Allen, D. D. (2004) Molecular modeling studies on the active binding site of the blood-brain barrier choline transporter. Bioorg Med Chem Lett 14, 3085-92.

(73) Griffith, D. A., and Jarvis, S. M. (1996) Nucleoside and nucleobase transport systems of mammalian cells. Biochim Biophys Acta 1286, 153-81.

251

(74) Cass, C. E., Young, J. D., and Baldwin, S. A. (1998) Recent advances in the molecular biology of nucleoside transporters of mammalian cells. Biochem Cell Biol 76, 761-70.

(75) Chandrasena, G., Giltay, R., Patil, S. D., Bakken, A., and Unadkat, J. D. (1997) Functional expression of human intestinal Na+-dependent and Na+- independent nucleoside transporters in Xenopus laevis oocytes. Biochem Pharmacol 53, 1909- 18.

(76) Patil, S. D., and Unadkat, J. D. (1997) Sodium-dependent nucleoside transport in the human intestinal brush- border membrane. Am J Physiol 272, G1314-20.

(77) Lai, Y., Bakken, A. H., and Unadkat, J. D. (2002) Simultaneous expression of hCNT1-CFP and hENT1-YFP in Madin-Darby canine kidney cells. Localization and vectorial transport studies. J Biol Chem 277, 37711-7.

(78) Jarvis, S. M., Thorn, J. A., and Glue, P. (1998) Ribavirin uptake by human erythrocytes and the involvement of nitrobenzylthioinosine-sensitive (es)- nucleoside transporters. Br J Pharmacol 123, 1587-92.

(79) Graham, K. A., Leithoff, J., Coe, I. R., Mowles, D., Mackey, J. R., Young, J. D., and Cass, C. E. (2000) Differential transport of cytosine-containing nucleosides by recombinant human concentrative nucleoside transporter protein hCNT1. Nucleosides, Nucleotides & Nucleic Acids 19, 415-34.

(80) Swaan, P. W., Koops, B. C., Moret, E. E., and Tukker, J. J. (1998) Mapping the binding site of the small intestinal peptide carrier (PepT1) using comparative molecular field analysis. Receptors Channels 6, 189-200.

(81) Viswanadhan, V. N., Ghose, A. K., and Weinstein, J. N. (1990) Mapping the binding site of the nucleoside transporter protein: a 3D-OSAR study. Biochimica et Biophysica Acta 1039, 356-66.

(82) Lum, P. Y., Ngo, L. Y., Bakken, A. H., and Unadkat, J. D. (2000) Human intestinal es nucleoside transporter: molecular characterization and nucleoside inhibitory profiles. Cancer Chemother Pharmacol 45, 273-8.

(83) Patil, S. D.]Ngo, L. Y., and Unadkat, J. D. (2000) Structure-inhibitory profiles of nucleosides for the human intestinal N1 and N2 Na+-nucleoside transporters. Cancer Chemother Pharmacol 46, 394-402.

(84) Ngo, L. Y., Patil, S. D., and Unadkat, J. D. (2001) Ontogenic and longitudinal activity of Na(+)-nucleoside transporters in the human intestine. Am J Physiol Gastrointest Liver Physiol 280, G475-81.

252

(85) Lang, T. T., Selner, M., Young, J. D., and Cass, C. E. (2001) Acquisition of human concentrative nucleoside transporter 2 (hcnt2) activity by gene transfer confers sensitivity to fluoropyrimidine nucleosides in drug-resistant leukemia cells. Molecular Pharmacology 60, 1143-52.

(86) Vickers, M. F., Kumar, R., Visser, F., Zhang, J., Charania, J., Raborn, R. T., Baldwin, S. A., Young, J. D., and Cass, C. E. (2002) Comparison of the interaction of uridine, cytidine, and other pyrimidine nucleoside analogues with recombinant human equilibrative nucleoside transporter 2 (hENT2) produced in Saccharomyces cerevisiae. Biochem Cell Biol 80, 639-44.

(87) Loewen, S. K., Ng, A. M., Yao, S. Y., Cass, C. E., Baldwin, S. A., and Young, J. D. (1999) Identification of amino acid residues responsible for the pyrimidine and purine nucleoside specificities of human concentrative Na(+) nucleoside cotransporters hCNT1 and hCNT2. J Biol Chem 274, 24475-84.

(88) Molina-Arcas, M., Bellosillo, B., Casado, F. J., Montserrat, E., Gil, J., Colomer, D., and Pastor-Anglada, M. (2003) Fludarabine uptake mechanisms in B-cell chronic lymphocytic leukemia. Blood 101, 2328-34.

(89) Zhang, J., Visser, F., Vickers, M. F., Lang, T., Robins, M. J., Nielsen, L. P., Nowak, I., Baldwin, S. A., Young, J. D., and Cass, C. E. (2003) Uridine binding motifs of human concentrative nucleoside transporters 1 and 3 produced in Saccharomyces cerevisiae. Mol Pharmacol 64, 1512-20.

(90) Huang, M., Wang, Y., Collins, M., Gu, J. J., Mitchell, B. S., and Graves, L. M. (2002) Inhibition of nucleoside transport by p38 MAPK inhibitors. J Biol Chem 277, 28364-7.

(91) Dean, P. M., Lloyd, D. G., and Todorov, N. P. (2004) De novo drug design: integration of structure-based and ligand-based methods. Curr Opin Drug Discov Devel 7, 347-53.

(92) Snyder, N. J., Tabas, L. B., Berry, D. M., Duckworth, D. C., Spry, D. O., and Dantzig, A. H. (1997) Structure-activity relationship of carbacephalosporins and cephalosporins: antibacterial activity and interaction with the intestinal proton- dependent dipeptide transport carrier of Caco-2 cells. Antimicrob Agents Chemother 41, 1649-57.

(93) Leibach, F. H., and Ganapathy, V. (1996) Peptide transporters in the intestine and the kidney. Annu Rev Nutr 16, 99-119.

(94) Han, H. K., Rhie, J. K., Oh, D. M., Saito, G., Hsu, C. P., Stewart, B. H., and Amidon, G. L. (1999) CHO/hPEPT1 cells overexpressing the human peptide transporter (hPEPT1) as an alternative in vitro model for peptidomimetic drugs. J Pharm Sci 88, 347-50.

253

(95) Zhang, E. Y., Fu, D. J., Pak, Y. A., Stewart, T., Mukhopadhyay, N., Wrighton, S. A., and Hillgren, K. M. (2004) Genetic Polymorphisms in Human Proton- Dependent Dipeptide Transporter PEPT1: Implications for the Functional Role of Pro586. J Pharmacol Exp Ther 310, 437-45.

(96) Brandsch, M., Knutter, I., and Leibach, F. H. (2004) The intestinal H+/peptide symporter PEPT1: structure-affinity relationships. Eur J Pharm Sci 21, 53-60.

(97) Yang, C. Y., Dantzig, A. H., and Pidgeon, C. (1999) Intestinal peptide transport systems and oral drug availability. Pharm Res 16, 1331-1343.

(98) Gomella, L., and Haist, S. (2004) Clinician's pocket drug reference 2004, McGraw-Hill.

(99) Covitz, K. M., Amidon, G. L., and Sadee, W. (1996) Human dipeptide transporter, hPEPT1, stably transfected into Chinese hamster ovary cells. Pharm Res 13, 1631-4.

(100) Li, J., and Hidalgo, I. J. (1996) Molecular modeling study of structural requirements for the oligopeptide transporter. J Drug Target 4, 9-17.

(101) Doring, F., Will, J., Amasheh, S., Clauss, W., Ahlbrecht, H., and Daniel, H. (1998) Minimal molecular determinants of substrates for recognition by the intestinal peptide transporter. J Biol Chem 273, 23211-23218.

(102) Ganapathy, M. E., Huang, W., Wang, H., Ganapathy, V., and Leibach, F. H. (1998) Valacyclovir: a substrate for the intestinal and renal peptide transporters PEPT1 and PEPT2. Biochem Biophys Res Commun 246, 470-5.

(103) Bailey, P. D., Boyd, C. A. R., Bronk, J. R., Collier, I. D., Meredith, D., Morgan, K. M., and Temple, C. S. (2000) How to make drugs orally active: a substrate template for peptide transporter PepT1. Angew Chem Int Ed 39, 506-508.

(104) Bretschneider, B., Brandsch, M., and Neubert, R. (1999) Intestinal transport of B- lactam antibiotics: analysis of the affinity at the H+/peptide symporter (PEPT1), the uptake into caco-2 cell monolayers and the transepithelial flux. Pharm Res 16, 55-61.

(105) Moore, V. A., Irwin, W. J., Timmins, P., Lambert, P. A., Chong, S., Dando, S. A., and Morrison, R. A. (2000) A rapid screening system to determine drug affinities for the intestinal dipeptide transporter 2: Affinities of ACE inhibitors. Int J Pharm 210, 29-44.

(106) Tse, F. L., Jaffe, J. M., and Troendle, A. (1992) Pharmacokinetics of fluvastatin after single and multiple doses in normal volunteers. J Clin Pharmacol 32, 630-8.

254

(107) Itagaki, S., Saito, Y., Kubo, S., Otsuka, Y., Yamamoto, Y., Kobayashi, M., Hirano, T., and Iseki, K. (2005) H+-dependent transport mechanism of nateglinide in the brush-border membrane of the rat intestine. J Pharmacol Exp Ther 312, 77-82.

(108) Ekins, S., and Swaan, P. W. (2004) Computational models for enzymes, transporters, channels and receptors relevant to ADME/TOX. Rev Comp Chem 20, 333-415.

(109) Koepsell, H. (2004) Polyspecific organic cation transporters: their functions and interactions with drugs. Trends Pharmacol Sci 25, 375-81.

(110) Koepsell, H., Schmitt, B. M., and Gorboulev, V. (2003) Organic cation transporters. Rev Physiol Biochem Pharmacol 150, 36-90.

(111) Jonker, J. W., and Schinkel, A. H. (2004) Pharmacological and physiological functions of the polyspecific organic cation transporters: OCT1, 2, and 3 (SLC22A1-3). J Pharmacol Exp Ther 308, 2-9.

(112) Wright, S. H., and Dantzler, W. H. (2004) Molecular and cellular physiology of renal organic cation and anion transport. Physiol Rev 84, 987-1049.

(113) Koepsell, H., and Endou, H. (2004) The SLC22 drug transporter family. Pflugers Arch 447, 666-76.

(114) Meijer, D. K., Mol, W. E., Muller, M., and Kurz, G. (1990) Carrier-mediated transport in the hepatic distribution and elimination of drugs, with special reference to the category of organic cations. J Pharmacokinet Biopharm 18, 35-70.

(115) Karbach, U., Kricke, J., Meyer-Wentrup, F., Gorboulev, V., Volk, C., Loffing- Cueni, D., Kaissling, B., Bachmann, S., and Koepsell, H. (2000) Localization of organic cation transporters OCT1 and OCT2 in rat kidney. Am J Physiol Renal Physiol 279, F679-87.

(116) van Montfoort, J. E., Muller, M., Groothuis, G. M., Meijer, D. K., Koepsell, H., and Meier, P. J. (2001) Comparison of "type I" and "type II" organic cation transport by organic cation transporters and organic anion-transporting polypeptides. J Pharmacol Exp Ther 298, 110-5.

(117) Kaewmokul, S., Chatsudthipong, V., Evans, K. K., Dantzler, W. H., and Wright, S. H. (2003) Functional mapping of rbOCT1 and rbOCT2 activity in the S2 segment of rabbit proximal tubule. Am J Physiol Renal Physiol 285, F1149-59.

(118) Jonker, J. W., Wagenaar, E., Van Eijl, S., and Schinkel, A. H. (2003) Deficiency in the organic cation transporters 1 and 2 (Oct1/Oct2 [Slc22a1/Slc22a2]) in mice abolishes renal secretion of organic cations. Mol Cell Biol 23, 7902-8.

255

(119) Motohashi, H., Sakurai, Y., Saito, H., Masuda, S., Urakami, Y., Goto, M., Fukatsu, A., Ogawa, O., and Inui, K. (2002) Gene expression levels and immunolocalization of organic ion transporters in the human kidney. J Am Soc Nephrol 13, 866-74.

(120) Ullrich, K. J. (1999) Affinity of drugs to the different renal transporters for organic anions and organic cations. Pharm Biotechnol 12, 159-79.

(121) Ullrich, K. J., Papavassiliou, F., David, C., Rumrich, G., and Fritzsch, G. (1991) Contraluminal transport of organic cations in the proximal tubule of the rat kidney. I. Kinetics of N1-methylnicotinamide and tetraethylammonium, influence of K+, HCO3-, pH; inhibition by aliphatic primary, secondary and tertiary amines and mono- and bisquaternary compounds. Pflugers Arch 419, 84-92.

(122) Wright, S. H., and Wunz, T. M. (1999) Influence of substrate structure on substrate binding to the renal organic cation/H+ exchanger. Pflugers Arch 437, 603-10.

(123) Barendt, W. M., and Wright, S. H. (2002) The human organic cation transporter (hOCT2) recognizes the degree of substrate ionization. J Biol Chem 277, 22491-6.

(124) Bednarczyk, D., Mash, E. A., Aavula, B. R., and Wright, S. H. (2000) NBD-TMA: a novel fluorescent substrate of the peritubular organic cation transporter of renal proximal tubules. Pflugers Arch 440, 184-92.

(125) Wright, S. H., Wunz, T. M., and Wunz, T. P. (1995) Structure and interaction of inhibitors with the TEA/H+ exchanger of rabbit renal brush border membranes. Pflugers Arch 429, 313-24.

(126) Busch, A. E., Karbach, U., Miska, D., Gorboulev, V., Akhoundova, A., Volk, C., Arndt, P., Ulzheimer, J. C., Sonders, M. S., Baumann, C., Waldegger, S., Lang, F., and Koepsell, H. (1998) Human neurons express the polyspecific cation transporter hOCT2, which translocates monoamine neurotransmitters, amantadine, and memantine. Mol Pharmacol 54, 342-52.

(127) Malo, C., and Berteloot, A. (1991) Analysis of kinetic data in transport studies: new insights from kinetic studies of Na(+)-D-glucose cotransport in human intestinal brush-border membrane vesicles using a fast sampling, rapid filtration apparatus. J Membr Biol 122, 127-41.

(128) Groves, C. E., Evans, K. K., Dantzler, W. H., and Wright, S. H. (1994) Peritubular organic cation transport in isolated rabbit proximal tubules. Am J Physiol 266, F450-8.

(129) Balant, L. P., and Gex-Fabry, M. (1990) Physiological pharmacokinetic modelling. Xenobiotica 20, 1241-57.

256

(130) Chaturvedi, P. R., Decker, C. J., and Odinecs, A. (2001) Prediction of pharmacokinetic properties using experimental approaches during early drug discovery. Curr Opin Chem Biol 5, 452-63.

(131) Ullrich, K. J., Rumrich, G., Neiteler, K., and Fritzsch, G. (1992) Contraluminal transport of organic cations in the proximal tubule of the rat kidney. II. Specificity: anilines, phenylalkylamines (catecholamines), heterocyclic compounds (pyridines, quinolines, acridines). Pflugers Arch 420, 29-38.

(132) Arndt, P., Volk, C., Gorboulev, V., Budiman, T., Popp, C., Ulzheimer-Teuber, I., Akhoundova, A., Koppatz, S., Bamberg, E., Nagel, G., and Koepsell, H. (2001) Interaction of cations, anions, and weak base quinine with rat renal cation transporter rOCT2 compared with rOCT1. Am J Physiol Renal Physiol 281, F454- 68.

(133) Urakami, Y., Okuda, M., Masuda, S., Akazawa, M., Saito, H., and Inui, K. (2001) Distinct characteristics of organic cation transporters, OCT1 and OCT2, in the basolateral membrane of renal tubules. Pharm Res 18, 1528-34.

(134) Green, R. E., Ricker, W. E., Attwood, W. L., Koh, Y. S., and Peters, L. (1959) Studies of the renal tubular transport characteristics of N1-methylnicotinamide and tetraalkylammonium compounds in the avian kidney. J Pharmacol Exp Ther 126, 195-201.

(135) Dresser, M. J., Xiao, G., Leabman, M. K., Gray, A. T., and Giacomini, K. M. (2002) Interactions of n-tetraalkylammonium compounds and biguanides with a human renal organic cation transporter (hOCT2). Pharm Res 19, 1244-7.

(136) Zhang, L., Gorset, W., Dresser, M. J., and Giacomini, K. M. (1999) The interaction of n-tetraalkylammonium compounds with a human organic cation transporter, hOCT1. J Pharmacol Exp Ther 288, 1192-8.

(137) Volk, C., Gorboulev, V., Budiman, T., Nagel, G., and Koepsell, H. (2003) Different affinities of inhibitors to the outwardly and inwardly directed substrate binding site of organic cation transporter 2. Mol Pharmacol 64, 1037-47.

(138) Gorboulev, V., Volk, C., Arndt, P., Akhoundova, A., and Koepsell, H. (1999) Selectivity of the polyspecific cation transporter rOCT1 is changed by mutation of aspartate 475 to glutamate. Mol Pharmacol 56, 1254-61.

(139) Leabman, M. K., Huang, C. C., Kawamoto, M., Johns, S. J., Stryke, D., Ferrin, T. E., DeYoung, J., Taylor, T., Clark, A. G., Herskowitz, I., and Giacomini, K. M. (2002) Polymorphisms in a human kidney xenobiotic transporter, OCT2, exhibit altered function. Pharmacogenetics 12, 395-405.

257

(140) Ekins, S. (2004) Predicting undesirable drug interactions with promiscuous proteins in silico. Drug Discovery Today.

(141) Tamai, I., Nezu, J., Uchino, H., Sai, Y., Oku, A., Shimane, M., and Tsuji, A. (2000) Molecular identification and characterization of novel members of the human organic anion transporter (OATP) family. Biochem Biophys Res Commun 273, 251-60.

(142) Bossuyt, X., Muller, M., Hagenbuch, B., and Meier, P. J. (1996) Polyspecific drug and steroid clearance by an organic anion transporter of mammalian liver. J Pharmacol Exp Ther 276, 891-896.

(143) Hagenbuch, B., and Meier, P. J. (2004) Organic anion transporting polypeptides of the OATP/ SLC21 family: phylogenetic classification as OATP/ SLCO superfamily, new nomenclature and molecular/functional properties. Pflugers Arch 447, 653-65.

(144) Satlin, L. M., Amin, V., and Wolkoff, A. W. (1997) Organic anion transporting polypeptide mediates organic anion/HCO3- exchange. J Biol Chem 272, 26340-5.

(145) Li, L., Lee, T. K., Meier, P. J., and Ballatori, N. (1998) Identification of glutathione as a driving force and leukotriene C4 as a substrate for oatp1, the hepatic sinusoidal organic solute transporter. J Biol Chem 273, 16184-91.

(146) Kim, R. B. (2003) Organic anion-transporting polypeptide (OATP) transporter family and drug disposition. Eur J Clin Invest 33, 1-5.

(147) Shitara, Y., Itoh, T., Sato, H., Li, A. P., and Sugiyama, Y. (2003) Inhibition of transporter-mediated hepatic uptake as a mechanism for drug-drug interaction between cerivastatin and cyclosporin A. J Pharmacol Exp Ther 304, 610-6.

(148) Shitara, Y., Hirano, M., Sato, H., and Sugiyama, Y. (2004) Gemfibrozil and its glucuronide inhibit the OATP2(OATP1B1: SLC21A6)-mediated hepatic uptake and CYP2C8-mediated metabolism of cerivastatin - Analysis of the mechanism of the clinically relevant drug-drug interaction between cerivastatin and gemfibrozil. J Pharmacol Exp Ther.

(149) Meier, P. J., and Stieger, B. (2000) Molecular Mechanisms in Bile Formation. News Physiol Sci 15, 89-93.

(150) Meier, P. J., Eckhardt, U., Schroeder, A., Hagenbuch, B., and Stieger, B. (1997) Substrate specificity of sinusoidal bile acid and organic anion uptake systems in rat and human liver. Hepatology 26, 1667-77.

(151) Ekins, S., Bravi, G., Binkley, S., Gillespie, J. S., Ring, B. J., Wikel, J. H., and Wrighton, S. A. (1999) Three dimensional-quantitative structure activity

258

relationship (3D-QSAR) analyses of inhibitors for CYP3A4. Journal of Pharmacology and Experimental Therapeutics 290, 429-438.

(152) Portmann, S., and Luthi, H. P. (2000) MOLEKEL: An Interactive Molecular Graphics Tool. Chimia 54, 766-770.

(153) Pang, K. S., Wang, P. J., Chung, A. Y., and Wolkoff, A. W. (1998) The modified dipeptide, enalapril, an angiotensin-converting enzyme inhibitor, is transported by the rat liver organic anion transport protein. Hepatology 28, 1341-6.

(154) Nozawa, T., Sugiura, S., Nakajima, M., Goto, A., Yokoi, T., Nezu, J., Tsuji, A., and Tamai, I. (2004) Involvement of organic anion transporting polypeptides in the transport of troglitazone sulfate: implications for understanding troglitazone hepatotoxicity. Drug Metab Dispos 32, 291-4.

(155) Cui, Y., Konig, J., Leier, I., Buchholz, U., and Keppler, D. (2001) Hepatic uptake of bilirubin and its conjugates by the human organic anion transporter SLC21A6. J Biol Chem 276, 9626-30.

(156) Yarim, M., Moro, S., Huber, R., Meier, P. J., Kaseda, C., Kashima, T., Hagenbuch, B., and Folkers, G. (2005) Application of QSAR analysis to organic anion transporting polypeptide 1a5 (Oatp1a5) substrates. Bioorg Med Chem 13, 463-71.

(157) Tirona, R. G., Leake, B. F., Wolkoff, A. W., and Kim, R. B. (2003) Human organic anion transporting polypeptide-C (SLC21A6) is a major determinant of the rifampin-mediated pregnane X receptor activiation. J Pharmacol Exp Ther 304, 223-228.

(158) Wang, P., Kim, R. B., Chowdhury, J. R., and Wolkoff, A. W. (2003) The human organic anion transport protein SLC21A6 is not sufficient for bilirubin transport. J Biol Chem 278, 20695-9.

(159) Briz, O., Serrano, M. A., Rebollo, N., Hagenbuch, B., Meier, P. J., Koepsell, H., and Marin, J. J. (2002) Carriers involved in targeting the cytostatic bile acid- cisplatin derivatives cis-diammine-chloro-cholylglycinate-platinum(II) and cis- diammine-bisursodeoxycholate-platinum(II) toward liver cells. Mol Pharmacol 61, 853-60.

(160) Briz, O., Serrano, M. A., MacIas, R. I., Gonzalez-Gallego, J., and Marin, J. J. (2003) Role of organic anion-transporting polypeptides, OATP-A, OATP-C and OATP-8, in the human placenta-maternal liver tandem excretory pathway for foetal bilirubin. Biochem J 371, 897-905.

259

(161) Meier-Abt, F., Faulstich, H., and Hagenbuch, B. (2004) Identification of phalloidin uptake systems of rat and human liver. Biochim Biophys Acta 1664, 64- 9.

(162) Fehrenbach, T., Cui, Y., Faulstich, H., and Keppler, D. (2003) Characterization of the transport of the bicyclic peptide phalloidin by human hepatic transport proteins. Naunyn Schmiedebergs Arch Pharmacol 368, 415-20.

(163) Tamai, I., Nozawa, T., Koshida, M., Nezu, J., Sai, Y., and Tsuji, A. (2001) Functional characterization of human organic anion transporting polypeptide B (OATP-B) in comparison with liver-specific OATP-C. Pharm Res 18, 1262-9.

(164) Sasaki, M., Suzuki, H., Ito, K., Abe, T., and Sugiyama, Y. (2002) Transcellular transport of organic anions across a double-transfected Madin-Darby canine kidney II cell monolayer expressing both human organic anion-transporting polypeptide (OATP2/SLC21A6) and Multidrug resistance-associated protein 2 (MRP2/ABCC2). J Biol Chem 277, 6497-503.

(165) Konig, J., Cui, Y., Nies, A. T., and Keppler, D. (2000) A novel human organic anion transporting polypeptide localized to the basolateral hepatocyte membrane. Am J Physiol Gastrointest Liver Physiol 278, G156-64.

(166) Nakai, D., Nakagomi, R., Furuta, Y., Tokui, T., Abe, T., Ikeda, T., and Nishimura, K. (2001) Human liver-specific organic anion transporter, LST-1, mediates uptake of pravastatin by human hepatocytes. J Pharmacol Exp Ther 297, 861-7.

(167) Hirano, M., Maeda, K., Shitara, Y., and Sugiyama, Y. (2004) Contribution of OATP2 (OATP1B1) and OATP8 (OATP1B3) to the Hepatic Uptake of Pitavastatin in Humans. J Pharmacol Exp Ther.

(168) Hsiang, B., Zhu, Y., Wang, Z., Wu, Y., Sasseville, V., Yang, W. P., and Kirchgessner, T. G. (1999) A novel human hepatic organic anion transporting polypeptide (OATP2). Identification of a liver-specific human organic anion transporting polypeptide and identification of rat and human hydroxymethylglutaryl-CoA reductase inhibitor transporters. J Biol Chem 274, 37161-8.

(169) Vavricka, S. R., Van Montfoort, J., Ha, H. R., Meier, P. J., and Fattinger, K. (2002) Interactions of rifamycin SV and rifampicin with organic anion uptake systems of human liver. Hepatology 36, 164-72.

(170) Schneck, D. W., Birmingham, B. K., Zalikowski, J. A., Mitchell, P. D., Wang, Y., Martin, P. D., Lasseter, K. C., Brown, C. D., Windass, A. S., and Raza, A. (2004) The effect of gemfibrozil on the pharmacokinetics of rosuvastatin. Clin Pharmacol Ther 75, 455-63.

260

(171) Abe, T., Kakyo, M., Tokui, T., Nakagomi, R., Nishio, T., Nakai, D., Nomura, H., Unno, M., Suzuki, M., Naitoh, T., Matsuno, S., and Yawo, H. (1999) Identification of a novel gene family encoding human liver-specific organic anion transporter LST-1. J Biol Chem 274, 17159-63.

(172) Eckhardt, U., Schroeder, A., Stieger, B., Hochli, M., Landmann, L., Tynes, R., Meier, P. J., and Hagenbuch, B. (1999) Polyspecific substrate uptake by the hepatic organic anion transporter Oatp1 in stably transfected CHO cells. Am J Physiol 276, G1037-42.

(173) Hata, S., Wang, P., Eftychiou, N., Ananthanarayanan, M., Batta, A., Salen, G., Pang, K. S., and Wolkoff, A. W. (2003) Substrate specificities of rat oatp1 and ntcp: implications for hepatic organic anion uptake. Am J Physiol Gastrointest Liver Physiol 285, G829-39.

(174) Kullack-Ublick, G.-A., Hagenbuch, B., Stieger, B., Wolkoff, A. W., and Meier, P. J. (1994) Functional characterization of the basolateral rat liver organic anion transporting polypeptide. Hepatology 20, 411-416.

(175) Shi, X., Bai, S., Ford, A. C., Burk, R. D., Jacquemin, E., Hagenbuch, B., Meier, P. J., and Wolkoff, A. W. (1995) Stable inducible expression of a functional rat liver organic anion transport protein in HeLa cells. J Biol Chem 270, 25591-5.

(176) Jacquemin, E., Hagenbuch, B., Stieger, B., Wolkoff, A. W., and Meier, P. J. (1994) Expression cloning of a rat liver Na(+)-independent organic anion transporter. Proc Natl Acad Sci U S A 91, 133-7.

(177) Reichel, C., Gao, B., Van Montfoort, J., Cattori, V., Rahner, C., Hagenbuch, B., Stieger, B., Kamisako, T., and Meier, P. J. (1999) Localization and function of the organic anion-transporting polypeptide Oatp2 in rat liver. Gastroenterology 117, 688-95.

(178) Eckhardt, U., Horz, J. A., Petzinger, E., Stuber, W., Reers, M., Dickneite, G., Daniel, H., Wagener, M., Hagenbuch, B., Stieger, B., and Meier, P. J. (1996) The peptide-based thrombin inhibitor CRC 220 is a new substrate of the basolateral rat liver organic anion-transporting polypeptide. Hepatology 24, 380-4.

(179) Gao, B., Hagenbuch, B., Kullak-Ublick, G. A., Benke, D., Aguzzi, A., and Meier, P. J. (2000) Organic anion-transporting polypeptides mediate transport of opioid peptides across blood-brain barrier. J Pharmacol Exp Ther 294, 73-9.

(180) Sugiyama, D., Kusuhara, H., Shitara, Y., Abe, T., Meier, P. J., Sekine, T., Endou, H., Suzuki, H., and Sugiyama, Y. (2001) Characterization of the efflux transport of 17b-estradiol-D-17b-glucuronide from the brain across the blood-brain barrier. J Pharmacol Exp Ther 298, 316-322.

261

(181) Shitara, Y., Sugiyama, D., Kusuhara, H., Kato, Y., Abe, T., Meier, P. J., Itoh, T., and Sugiyama, Y. (2002) Comparative inhibitory effects of different compounds on rat oatpl (slc21a1)- and Oatp2 (Slc21a5)-mediated transport. Pharm Res 19, 147-53.

(182) Kanai, N., Lu, R., Bao, Y., Wolkoff, A. W., Vore, M., and Schuster, V. L. (1996) Estradiol 17 beta-D-glucuronide is a high-affinity substrate for oatp organic anion transporter. Am J Physiol 270, F326-31.

(183) Kouzuki, H., Suzuki, H., Ito, K., Ohashi, R., and Sugiyama, Y. (1999) Contribution of organic anion transporting polypeptide to uptake of its possible substrates into rat hepatocytes. J Pharmacol Exp Ther 288, 627-34.

(184) Ishizuka, H., Konno, K., Naganuma, H., Nishimura, K., Kouzuki, H., Suzuki, H., Stieger, B., Meier, P. J., and Sugiyama, Y. (1998) Transport of temocaprilat into rat hepatocytes: role of organic anion transporting polypeptide. J Pharmacol Exp Ther 287, 37-42.

(185) Cvetkovic, M., Leake, B., Fromm, M. F., Wilkinson, G. R., and Kim, R. B. (1999) OATP and P-glycoprotein transporters mediate the cellular uptake and excretion of fexofenadine. Drug Metab Dispos 27, 866-71.

(186) van Montfoort, J. E., Stieger, B., Meijer, D. K., Weinmann, H. J., Meier, P. J., and Fattinger, K. E. (1999) Hepatic uptake of the magnetic resonance imaging contrast agent gadoxetate by the organic anion transporting polypeptide Oatp1. J Pharmacol Exp Ther 290, 153-7.

(187) Kontaxi, M., Echkardt, U., Hagenbuch, B., Stieger, B., Meier, P. J., and Petzinger, E. (1996) Uptake of the mycotoxin ochratoxin A in liver cells occurs via the cloned organic anion transporting polypeptide. J Pharmacol Exp Ther 279, 1507- 13.

(188) Hagenbuch, B., Lubbert, H., Stieger, B., and Meier, P. J. (1990) Expression of the hepatocyte Na+/bile acid cotransporter in Xenopus laevis oocytes. J Biol Chem 265, 5357-60.

(189) Meng, L. J., Wang, P., Wolkoff, A. W., Kim, R. B., Tirona, R. G., Hofmann, A. F., and Pang, K. S. (2002) Transport of the sulfated, amidated bile acid, sulfolithocholyltaurine, into rat hepatocytes is mediated by Oatp1 and Oatp2. Hepatology 35, 1031-40.

(190) Meijer, O. C., de Lange, E. C., Breimer, D. D., de Boer, A. G., Workel, J. O., and de Kloet, E. R. (1998) Penetration of dexamethasone into brain glucocorticoid targets is enhanced in mdr1A P-glycoprotein knockout mice. Endocrinology 139, 1789-93.

262

(191) Karssen, A. M., Meijer, O. C., van der Sandt, I. C., Lucassen, P. J., de Lange, E. C., de Boer, A. G., and de Kloet, E. R. (2001) Multidrug resistance P-glycoprotein hampers the access of cortisol but not of corticosterone to mouse and human brain. Endocrinology 142, 2686-94.

(192) Nakayama, A., Eguchi, O., Hatakeyama, M., Saitoh, H., and Takada, M. (1999) Different absorption behaviors among steroid hormones due to possible interaction with P-glycoprotein in the rat small intestine. Biol Pharm Bull 22, 535- 8.

(193) Bourgeois, S., Gruol, D. J., Newby, R. F., and Rajah, F. M. (1993) Expression of an mdr gene is associated with a new form of resistance to dexamethasone- induced apoptosis. Mol Endocrinol 7, 840-51.

(194) Gruol, D. J., and Bourgeois, S. (1997) Chemosensitizing steroids: glucocorticoid receptor agonists capable of inhibiting P-glycoprotein function. Cancer Res 57, 720-7.

(195) Ueda, K., Okamura, N., Hirai, M., Tanigawara, Y., Saeki, T., Kioka, N., Komano, T., and Hori, R. (1992) Human P-glycoprotein transports cortisol, aldosterone, and dexamethasone, but not progesterone. J Biol Chem 267, 24248-52.

(196) van Kalken, C. K., Broxterman, H. J., Pinedo, H. M., Feller, N., Dekker, H., Lankelma, J., and Giaccone, G. (1993) Cortisol is transported by the multidrug resistance gene product P-glycoprotein. Br J Cancer 67, 284-9.

(197) Barnes, K. M., Dickstein, B., Cutler, G. B., Jr., Fojo, T., and Bates, S. E. (1996) Steroid treatment, accumulation, and antagonism of P-glycoprotein in multidrug- resistant cells. Biochemistry 35, 4820-7.

(198) Schinkel, A. H., Wagenaar, E., van Deemter, L., Mol, C. A., and Borst, P. (1995) Absence of the mdr1a P-Glycoprotein in mice affects tissue distribution and pharmacokinetics of dexamethasone, digoxin, and cyclosporin A. J Clin Invest 96, 1698-705.

(199) Hsu, S. I., Lothstein, L., and Horwitz, S. B. (1989) Differential overexpression of three mdr gene family members in multidrug-resistant J774.2 mouse cells. Evidence that distinct P-glycoprotein precursors are encoded by unique mdr genes. J Biol Chem 264, 12053-62.

(200) Homolya, L., Hollo, Z., Germann, U. A., Pastan, I., Gottesman, M. M., and Sarkadi, B. (1993) Fluorescent cellular indicators are extruded by the multidrug resistance protein. J Biol Chem 268, 21493-6.

(201) Ford, J. M. (1996) Experimental reversal of P-glycoprotein-mediated multidrug resistance by pharmacological chemosensitisers. Eur J Cancer 32A, 991-1001.

263

(202) Yates, C. R., Vysokanov, A., Mukherjee, A., Ludden, T. M., Tolley, E., Meduri, G. U., and Dalton, J. T. (2001) Time-variant increase in methylprednisolone clearance in patients with acute respiratory distress syndrome: a population pharmacokinetic study. J Clin Pharmacol 41, 415-24.

(203) Torok, M., Gutmann, H., Fricker, G., and Drewe, J. (1999) Sister of P- glycoprotein expression in different tissues. Biochem Pharmacol 57, 833-5.

(204) Raggers, R. J., van Helvoort, A., Evers, R., and van Meer, G. (1999) The human multidrug resistance protein MRP1 translocates sphingolipid analogs across the plasma membrane. J Cell Sci 112 (Pt 3), 415-22.

(205) Wils, P., Warnery, A., Phung-Ba, V., Legrain, S., and Scherman, D. (1994) High lipophilicity decreases drug transport across intestinal epithelial cells. J Pharmacol Exp Ther 269, 654-8.

(206) Rothblat, G. H., and Phillips, M. C. (1982) Mechanism of cholesterol efflux from cells. Effects of acceptor structure and concentration. J Biol Chem 257, 4775-82.

(207) Luxon, B. A., and Weisiger, R. A. (1992) A new method for quantitating intracellular transport: application to the thyroid hormone 3,5,3'-triiodothyronine. Am J Physiol 263, G733-41.

(208) Gruol, D. J., Vo, Q. D., and Zee, M. C. (1999) Profound differences in the transport of steroids by two mouse P-glycoproteins. Biochem Pharmacol 58, 1191-9.

(209) Saitoh, H., Hatakeyama, M., Eguchi, O., Oda, M., and Takada, M. (1998) Involvement of intestinal P-glycoprotein in the restricted absorption of methylprednisolone from rat small intestine. J Pharm Sci 87, 73-5.

(210) Koszdin, K. L., Shen, D. D., and Bernards, C. M. (2000) Spinal cord bioavailability of methylprednisolone after intravenous and intrathecal administration: the role of P-glycoprotein. Anesthesiology 92, 156-63.

(211) Ayesh, S., Shao, Y.-M., and Stein, W. D. (1996) Co-operative, competitive and non-competitive interactions between modulators of P-glycoprotein. Biochimica et Biophysica Acta 1316, 8-18.

(212) Dey, S., Ramachandra, M., Pastan, I., Gottesman, M. M., and Ambudkar, S. (1997) Evidence for two nonidentical drug-interaction sites in the human P- glycoprotein. Proc Natl Acad Sci USA 94, 10594-10599.

(213) Scala, S., Akhmed, N., Rao, U. S., Paull, K., Lan, L.-B., Dickstein, B., Lee, J.-S., Elgemeie, G. H., Stein, W. D., and Bates, S. E. (1997) P-glycoprotein substrates and antagonists cluster into two distinct groups. Mol Pharmacol 51, 1024-1033.

264

(214) Shapiro, A. B., Fox, K., Lam, P., and Ling, V. (1999) Stimulation of P- glycoprotein-mediated drug transport by prazosin and progersterone. Evidence for a third site. Eur J Biochem 259, 841-850.

(215) Shapiro, A. B., and Ling, V. (1997) Positively cooperative sites for drug transport by P-glycoprotein with distinct drug specificities. Eur J Biochem 250, 130-137.

(216) Pleban, K., Kopp, S., Csaszar, E., Peer, M., Hrebicek, T., Rizzi, A., Ecker, G. F., and Chiba, P. (2005) P-glycoprotein substrate binding domains are located at the transmembrane domain/transmembrane domain interfaces: a combined photoaffinity labeling-protein homology modeling approach. Mol Pharmacol 67, 365-74.

(217) Rosenberg, M. F., Kamis, A. B., Callaghan, R., Higgins, C. F., and Ford, R. C. (2003) Three-dimensional structures of the mammalian multidrug resistance P- glycoprotein demonstrate major conformational changes in the transmembrane domains upon nucleotide binding. J Biol Chem 278, 8294-9.

(218) Stouch, T. R., and Gudmundsson, O. (2002) Progress in understanding the structure-activity relationships of P-glycoprotein. Adv Drug Deliv Rev 54, 315-28.

(219) Pajeva, I. K., and Wiese, M. (2002) Pharmacophore model of drugs involved in P- glycoprotein multidrug resistance: explanation of structural variety (hypothesis). J Med Chem.

(220) Garrigues, A., Loiseau, N., Delaforge, M., Ferte, J., Garrigos, M., Andre, F., and Orlowski, S. (2002) Characterization of two pharmacophores on the multidrug transporter P-glycoprotein. Mol Pharmacol 62, 1288-1298.

(221) Langer, T., Eder, M., Hoffmann, R. D., Chiba, P., and Ecker, G. F. (2004) Lead identification for modulators of multidrug resistance based on in silico screening with a pharmacophoric feature model. Arch Pharm (Weinheim) 337, 317-27.

(222) Penzotti, J. E., Lamb, M. L., Evensen, E., and Grootenhuis, P. D. (2002) A computational ensemble pharmacophore model for identifying substrates of P- glycoprotein. J Med Chem 45, 1737-40.

(223) Güner, O. F. (2000) Pharmacophore perception, development, and use in drug design, International University Line, LaJolla, CA.

(224) Pajeva, I. K., Globisch, C., and Wiese, M. (2004) Structure-function relationships of multidrug resistance P-glycoprotein. J Med Chem 47, 2523-33.

(225) Szakacs, G., Annereau, J. P., Lababidi, S., Shankavaram, U., Arciello, A., Bussey, K. J., Reinhold, W., Guo, Y., Kruh, G. D., Reimers, M., Weinstein, J. N., and

265

Gottesman, M. M. (2004) Predicting drug sensitivity and resistance: profiling ABC transporter genes in cancer cells. Cancer Cell 6, 129-37.

(226) Clement, O. O., Freeman, C. M., Hartmann, R. W., Handratta, V. D., Vasaitis, T. S., Brodie, A. M., and Njar, V. C. (2003) Three dimensional pharmacophore modeling of human CYP17 inhibitors. Potential agents for prostate cancer therapy. J Med Chem 46, 2345-51.

(227) Kahnberg, P., Howard, M. H., Liljefors, T., Nielsen, M., Nielsen, E. O., Sterner, O., and Pettersson, I. (2004) The use of a pharmacophore model for identification of novel ligands for the benzodiazepine binding site of the GABAA receptor. J Mol Graph Model 23, 253-61.

(228) Koide, Y., Hasegawa, T., Takahashi, A., Endo, A., Mochizuki, N., Nakagawa, M., and Nishida, A. (2002) Development of novel EDG3 antagonists using a 3D database search and their structure-activity relationships. J Med Chem 45, 4629- 38.

(229) Kurogi, Y., and Guner, O. F. (2001) Pharmacophore modeling and three- dimensional database searching for drug design using catalyst. Curr Med Chem 8, 1035-1055.

(230) Dalton, J. T., Mukherjee, A., Zhu, Z., Kirkovsky, L., and Miller, D. D. (1998) Discovery of nonsteroidal androgens. Biochem Biophys Res Commun 244, 1-4.

(231) Yin, D., Xu, H., He, Y., Kirkovsky, L. I., Miller, D. D., and Dalton, J. T. (2003) Pharmacology, pharmacokinetics, and metabolism of acetothiolutamide, a novel nonsteroidal agonist for the androgen receptor. J Pharmacol Exp Ther 304, 1323- 33.

(232) Negro-Vilar, A. (1999) Selective androgen receptor modulators (SARMs): a novel approach to androgen therapy for the new millennium. J Clin Endocrinol Metab 84, 3459-62.

(233) Yin, D., He, Y., Perera, M. A., Hong, S. S., Marhefka, C., Stourman, N., Kirkovsky, L., Miller, D. D., and Dalton, J. T. (2003) Key structural features of nonsteroidal ligands for binding and activation of the androgen receptor. Mol Pharmacol 63, 211-23.

(234) Yin, D., Gao, W., Kearbey, J. D., Xu, H., Chung, K., He, Y., Marhefka, C. A., Veverka, K. A., Miller, D. D., and Dalton, J. T. (2003) Pharmacodynamics of selective androgen receptor modulators. J Pharmacol Exp Ther 304, 1334-40.

(235) Mukherjee, A., Kirkovsky, L. I., Kimura, Y., Marvel, M. M., Miller, D. D., and Dalton, J. T. (1999) Affinity labeling of the androgen receptor with nonsteroidal chemoaffinity ligands. Biochem Pharmacol 58, 1259-67.

266

(236) He, Y., Yin, D., Perera, M., Kirkovsky, L., Stourman, N., Li, W., Dalton, J. T., and Miller, D. D. (2002) Novel nonsteroidal ligands with high binding affinity and potent functional activity for the androgen receptor. Eur J Med Chem 37, 619-34.

(237) Waller, C. L., Juma, B. W., Gray, L. E., Jr., and Kelce, W. R. (1996) Three- dimensional quantitative structure--activity relationships for androgen receptor ligands. Toxicol Appl Pharmacol 137, 219-27.

(238) Gohlke, H., and Klebe, G. (2002) DrugScore meets CoMFA: adaptation of fields for molecular comparison (AFMoC) or how to tailor knowledge-based pair- potentials to a particular protein. J Med Chem 45, 4153-70.

(239) Poujol, N., Wurtz, J. M., Tahiri, B., Lumbroso, S., Nicolas, J. C., Moras, D., and Sultan, C. (2000) Specific recognition of androgens by their nuclear receptor. A structure-function study. Journal of Biological Chemistry 275, 24022-31.

(240) Marhefka, C. A., Moore, B. M., 2nd, Bishop, T. C., Kirkovsky, L., Mukherjee, A., Dalton, J. T., and Miller, D. D. (2001) Homology modeling using multiple molecular dynamics simulations and docking studies of the human androgen receptor ligand binding domain bound to testosterone and nonsteroidal ligands. J Med Chem 44, 1729-40.

(241) Zhi, L., Tegley, C. M., Marschke, K. B., and Jones, T. K. (1999) Switching androgen receptor antagonists to agonists by modifying C-ring substituents on piperidino[3,2-g]quinolinone. Bioorg Med Chem Lett 9, 1009-12.

(242) Clark, M., Cramer III, R. D., Jones, D. M., Patterson, D. E., and Simeroth, P. E. (1990) Comparative Molecular Field Analysis (CoMFA) 2. Toward the use with 3D structural database. Tetrahedron Comput. Methodol. 3, 47-59.

(243) Rarey, M., Kramer, B., Lengauer, T., and Klebe, G. (1996) A fast flexible docking method using an incremental construction algorithm. J Mol Biol 261, 470-89.

(244) Matias, P. M., Donner, P., Coelho, R., Thomaz, M., Peixoto, C., Macedo, S., Otto, N., Joschko, S., Scholz, P., Wegg, A., Basler, S., Schafer, M., Egner, U., and Carrondo, M. A. (2000) Structural evidence for ligand specificity in the binding domain of the human androgen receptor. Implications for pathogenic gene mutations. J Biol Chem 275, 26164-71.

(245) Veldscholte, J., Berrevoets, C. A., Ris-Stalpers, C., Kuiper, G. G., Jenster, G., Trapman, J., Brinkmann, A. O., and Mulder, E. (1992) The androgen receptor in LNCaP cells contains a mutation in the ligand binding domain which affects steroid binding characteristics and response to antiandrogens. J Steroid Biochem Mol Biol 41, 665-9.

267

(246) Batch, J. A., Williams, D. M., Davies, H. R., Brown, B. D., Evans, B. A., Hughes, I. A., and Patterson, M. N. (1992) Androgen receptor gene mutations identified by SSCP in fourteen subjects with androgen insensitivity syndrome. Human Molecular Genetics 1, 497-503.

(247) Chawla, A., Repa, J. J., Evans, R. M., and Mangelsdorf, D. J. (2001) Nuclear receptors and lipid physiology: opening the X-files. Science 294, 1866-70.

(248) Goodwin, B., and Kliewer, S. A. (2002) Nuclear receptors. I. Nuclear receptors and bile acid homeostasis. Am J Physiol Gastrointest Liver Physiol 282, G926-31.

(249) Evans, R. M. (1988) The steroid and thyroid hormone receptor superfamily. Science 240, 889-95.

(250) Ekins, S., Mirny, L., and Schuetz, E. G. (2002) A ligand-based approach to understanding selectivity of nuclear hormone receptors PXR, CAR, FXR, LXRa and LXRb. Pharm Res 19, 1788-1800.

(251) Godzik, A., and Skolnick, J. (1992) Sequence-structure matching in globular proteins: application to supersecondary and tertiary structure determination. Proc Natl Acad Sci U S A 89, 12098-102.

(252) Afzelius, L., Zamora, I., Ridderstrom, M., Andersson, T. B., Karlen, A., and Masimirembwa, C. M. (2001) Competitive CYP2C9 inhibitors: enzyme inhibition studies, protein homology modeling, and three dimensional quantitative structure activity relationship analysis. Mol Pharmacol 59, 909-919.

(253) Mi, L. Z., Devarakonda, S., Harp, J. M., Han, Q., Pellicciari, R., Willson, T. M., Khorasanizadeh, S., and Rastinejad, F. (2003) Structural basis for bile acid binding and activation of the nuclear receptor FXR. Mol Cell 11, 1093-100.

(254) Downes, M., Verdecia, M. A., Roecker, A. J., Hughes, R., Hogenesch, J. B., Kast- Woelbern, H. R., Bowman, M. E., Ferrer, J. L., Anisfeld, A. M., Edwards, P. A., Rosenfeld, J. M., Alvarez, J. G., Noel, J. P., Nicolaou, K. C., and Evans, R. M. (2003) A chemical, genetic, and structural analysis of the nuclear bile acid receptor FXR. Mol Cell 11, 1079-92.

(255) Shimada, J., Ekins, S., Elkin, C., Shaknovich, E. I., and Wery, J.-P. (2002) Integrating computer-based de novo drug design and multidimensional filtering for desirable drugs. Targets 1, 196-205.

(256) Ekins, S., Andreyev, S., Ryabov, A., Kirilov, E., Rakhmatulin, E. A., Bugrim, A., and Nikolskaya, T. (2005) Computational Prediction of Human Drug Metabolism. Exp Opin Drug Metab Toxicol in press.

268

(257) Ekins, S., Nikolsky, Y., and Nikolskaya, T. (2005) Techniques: Application of Systems Biology to Absorption, Distribution, Metabolism, Excretion, and Toxicity. Trends Pharmacol Sci in press.

269

INDEX

Dopamine transporter ...... 22, 23, 30, 32 A

G ABC transporter ...... 10

Apical sodium-dependent bile acid transporter25, GASP...... 13, 173

28, 30, 32 Glucocorticoids ...... 144, 149, 154

Glut1...... 6 B GOLPE 17, 21, 39, 41, 44, 45, 47, 53, 59, 60, 61, Bile acid receptor ...... 225 239 C H CoMFA11, 16, 17, 19, 20, 21, 22, 23, 25, 28, 31, HIPHOP ...14, 65, 68, 76, 77, 173, 175, 180, 239 32, 35, 37, 39, 40, 41, 44, 45, 47, 49, 50, 53, HypoGen ...... 12, 15, 243 54, 59, 60, 61, 62, 69, 83, 84, 92, 93, 96, 97, M 100, 109, 110, 114, 125, 152, 153, 156, 157,

160, 170, 171, 195, 196, 198, 199, 201, 202, Modeller ...... 9, 30, 226

204, 205, 207, 208, 209, 210, 218, 219, 221, N 222, 239, 240, 241 Norepinephrine transporter...... 22, 23, 30 CoMSIA ....11, 17, 20, 31, 32, 69, 153, 156, 157, Nucleoside transporter...... 26, 33 161 O D OATP17, 115, 116, 118, 120, 123, 125, 135, 138, DISCO.....11, 12, 14, 31, 32, 36, 38, 47, 58, 152, 240 155, 156, 160, 170, 239

270

P 158, 159, 160, 161, 169, 172, 174, 175, 176,

177, 178, 179, 180, 181, 182, 185, 193, 241 p38 MAPK inhibitors...... 48, 49, 50, 62

Peptide transporter.63, 64, 65, 67, 68, 69, 71, 74, S

76, 77, 118, 180, 181, 182, 239 SARM...... 194 P-gpiii, vii, 3, 7, 8, 9, 26, 30, 31, 34, 71, 100, 115, Serotonin transporter ...... 22, 23, 30, 32 129, 144, 145, 147, 148, 149, 153, 154, 155,

271