ACTAUNIVERSITATIS

UPSALIENSISUPPSALA

2020

Digital Comprehensive Summaries of Uppsala Dissertationsfrom the Faculty of Science and Technology 1889

New Paradigms in GPCR DrugDiscovery

Structure Prediction and Design of Ligands withTailored Properties

MARIAMA JAITEH

ISSN 1651-6214ISBN 978-91-513-0836-4urn:nbn:se:uu:diva-399133

Dissertation presented at Uppsala University to be publicly examined in Room A1:111a,BMC, Husargatan 3, Uppsala, Friday, 14 February 2020 at 09:15 for the degree ofDoctor of Philosophy. The examination will be conducted in English. Faculty examiner:Professor David E Gloriam (University of Copenhagen, Department of Drug Design andPharmacology).

AbstractJaiteh, M. 2020. New Paradigms in GPCR Drug Discovery. Structure Prediction andDesign of Ligands with Tailored Properties. Digital Comprehensive Summaries of UppsalaDissertations from the Faculty of Science and Technology 1889. 63 pp. Uppsala: ActaUniversitatis Upsaliensis. ISBN 978-91-513-0836-4.

G protein-coupled receptors (GPCRs) constitute a large superfamily of membrane proteinswith key roles in cellular signaling. Upon activation by a ligand, GPCRs transduce signalsfrom the extracellular to the intracellular environment. GPCRs are important drug targets andare associated with diseases such as central nervous system (CNS) disorders, cardiovasculardiseases, cancer, and diabetes. Currently, 34% of FDA-approved drugs mediate their effects viamodulation of GPCRs. Research during the past decades has resulted in a deeper understandingof GPCR structure and function. Moreover, recent breakthroughs in structural biology allowedthe determination of several atomic resolution GPCR structures. New paradigms in GPCRpharmacology have also emerged that can lead to improved drugs. Together, these advancesprovide new avenues for structure-based drug discovery. The work in this thesis focused on howthe large amount of structural data gathered over the last decades can be used to model GPCRtargets for which no experimental structures are available, and the use of structure-based virtualscreening (SBVS) campaigns to identify ligands with tailored pharmacological properties. Inpaper I, we investigated how template selection affects the virtual screening performance ofhomology models of the D2 dopamine receptor (D2R) and serotonin 5-HT2A receptor (5-HT2AR).In papers II and III, SBVS methods were used to identify dual inhibitors of the A2A adenosinereceptor (A2AAR) and an enzyme, which could be relevant for treatment of Parkinson’s Disease,and functionally selective D2R ligands from a focused library. Finally, we also investigated howstructural information can complement computational and biophysical methods to model andcharacterize the A2AAR-D2R heterodimer (paper IV).

Keywords: G Protein-Coupled Receptor, Molecular Docking, Virtual Screening, HomologyModeling, Molecular Dynamics Simulation, Chemical Library, Functionally Selective Ligand,Polypharmacology, Dimerization

Mariama Jaiteh, Department of Cell and Molecular Biology, Computational Biology andBioinformatics, Box 596, Uppsala University, SE-751 24 Uppsala, Sweden.

© Mariama Jaiteh 2020

ISSN 1651-6214ISBN 978-91-513-0836-4urn:nbn:se:uu:diva-399133 (http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-399133)

To Isatou

List of papers

This thesis is based on the following papers, which are referred to in the textby their Roman numerals.

I Jaiteh, M., Rodríguez-Espigares, I., Selent, J., and Carlsson, J.Performance of Virtual Screening against GPCR Homology Models:Impact of Template Selection and Treatment of Binding Site Plasticity.PLoS Comput Biol. Manuscript under review.

II Jaiteh, M., Zeifman, A., Saarinen, M., Svenningsson, P., Bréa, J.,Loza, M. I., and Carlsson, J. (2018) Docking Screens for DualInhibitors of Disparate Drug Targets for Parkinson’s Disease. J. Med.Chem. 61, 5269–5278.

III Männel, B., Jaiteh, M., Zeifman, A., Randakova, A., Möller, A.,Hübner, A. Gmeiner, P., and Carlsson, J. (2017) Structure-GuidedScreening for Functionally Selective D2 Dopamine Receptor Ligandsfrom a Virtual Chemical Library. ACS Chem. Biol. 12, 2652–2661.

IV Borroto-Escuela, D. O., Rodriguez, D., Romero-Fernandez, W., Kapla,J., Jaiteh, M., Ranganathan, A., Lazarova, T., Fuxe, K., and Carlsson,J. (2018) Mapping the Interface of a GPCR Dimer: A Structural Modelof the A2A Adenosine and D2 Dopamine Receptor Heteromer. Front.Pharmacol. 9, 829.

The following papers were not included in this thesis.

V Matricon, P., Ranganathan, A., Warnick, E., Gao, Z-G., Rudling, A.,Lambertucci, C., Marucci, G., Ezzati, A., Jaiteh, M., Dal Ben, D.,Jacobson, K. A., and Carlsson, J. (2017) Fragment Optimization forGPCRs by Molecular Dynamics Free Energy Calculations: ProbingDruggable Subpockets of the A2A Adenosine Receptor Binding Site.Sci. Rep. 7, 6398.

VI Jaiteh, M., Taly, A., and Hénin, J. (2016) Evolution of PentamericLigand-Gated Ion Channels: Pro-Loop Receptors. PLoS One 11,e0151934.

Reprints were made with permission from the publishers.

Contents

List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.1 The receptor concept and birth of the G protein-coupled

receptor superfamily . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.2 GPCR superfamily . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.3 GPCR signaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.3.1 The canonical G protein pathway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.3.2 Non-canonical pathways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.4 GPCR pharmacology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.4.1 Orthosteric ligands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.4.2 Allosteric modulation and GPCR oligomerization . . . . . . . 14

1.5 GPCR structural biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151.6 New paradigms in GPCR drug discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Computational structure prediction of GPCRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.1 Ab initio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.2 Threading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.3 Homology modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.3.1 Template selection and sequence alignment . . . . . . . . . . . . . . . . . . 192.3.2 Build conserved and non-conserved regions . . . . . . . . . . . . . . . . . 202.3.3 Model evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.3.4 Homology modeling with MODELLER . . . . . . . . . . . . . . . . . . . . . . . . 22

2.4 Protein-protein docking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4.1 The sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4.2 The scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4.3 Protein-protein docking with HADDOCK . . . . . . . . . . . . . . . . . . . . 23

3 Molecular dynamics simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243.1 The force field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.1.1 The bonded interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243.1.2 The non-bonded interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2 The simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263.2.1 Initial parameters and settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263.2.2 Equations of motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4 Structure-based virtual screening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284.1 Molecular docking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.1.1 Receptor binding site preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.1.2 Library preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284.1.3 The molecular docking process: sampling and scoring 294.1.4 Retrospective screening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294.1.5 Molecular docking with DOCK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5 Paper I: Performance of Virtual Screening against GPCR HomologyModels: Impact of Template Selection and Treatment of Binding SitePlasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

6 Paper II: Docking Screens for Dual Inhibitors of Disparate DrugTargets for Parkinson’s Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

7 Paper III: Structure-Guided Screening for Functionally Selective D2

Dopamine Receptor Ligands from a Virtual Chemical Library . . . . . . . . . . . . . 39

8 Paper IV: Mapping the Interface of a GPCR Dimer: A StructuralModel of A2A Adenosine and D2 Dopamine Receptor Heteromer . . . . . . . . 43

9 Conclusions and perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

10 Sammanfattning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

11 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

List of Abbreviations

3D Three-dimensional5-HT2AR Serotonin 5-HT2A Receptor7TM Seven TransmembraneA2AAR A2A Adenosine ReceptorAUC Area Under the CurveBS Binding SiteCAPRI Critical Assessment of Predicted InteractionsCASP Critical Assessment of Techniques for Protein

Structure PredictionCNS Central Nervous Systemcryo-EM Cryogenic Electron MicroscopyD2R D2 Dopamine ReceptorDOPE Discrete Optimized Protein EnergyECL Extracellular LoopGPCR G Protein-Coupled ReceptorHM Homology ModelingHTS High-Throughput ScreeningICL Intracellular LoopMAO-B Monoamine Oxidase BMD Molecular DynamicsMSA Multiple Sequence AlignmentMW Molecular WeightNMR Nuclear Magnetic ResonancePBC Periodic Boundary ConditionsPD Parkinson’s DiseasePDB Protein Data BankRMSD Root Mean Square DeviationROC Receiver Operating CharacteristicSBVS Structure-Based Virtual ScreeningTM TransmembraneVS Virtual Screening

9

1. Introduction

1.1 The receptor concept and birth of the Gprotein-coupled receptor superfamily

The idea of receptors was highly controversial when first introduced by J. N.Langley in 1905. Langley described the existence of ”receptive substanceswhich are acted upon by chemical bodies and in certain cases by nervous stim-uli. The receptive substance affects or is capable of affecting the metabolismof the chief substance” [1]. A few years later, Ehrlich proposed the “lock andkey” theory to describe the specific interaction between receptor and ligand [2].These were the first assertions on the existence of a receptor that binds specifi-cally to a ligand to trigger cellular responses. After eight decades of resistance,the pioneering work of Robert Lefkowitz’s group and advances in biochemi-cal techniques (e.g. radioligand binding and purification) led to the identifica-tion of the first G protein-coupled receptor (GPCR), the β-adrenergic receptor.Cloning of the β2-adrenergic receptor [3] led to the discovery of a protein withseven transmembrane (7TM) spanning helices, which had been previously ob-served for rhodopsin and bacteriorhodopsin. Thereafter, the number of clonedreceptors exhibiting the 7TM helix topology grew rapidly, giving birth to theGPCR superfamily [4].

1.2 GPCR superfamilyComprising more than 800 members, the GPCRs constitute one of the largestsuperfamilies of membrane proteins in the human genome. Based on phyloge-netic studies, they are classified into five main subfamilies [5]: the Rhodopsin-like (also known as class A, which accounts for ∼87% of human GPCRs),the Secretin (class B, 15 members), the Glutamate (class C, 15 members),the Adhesion (24 members), and the Frizzled/Taste2 (24 members) receptors.They are characterized by an N-terminal segment, seven transmembrane (TM)helices connected by three extracellular loops (ECLs) and three intracellularloops (ICLs), and a C-terminal segment. Because of their characteristic topol-ogy, the GPCRs are called the “7TM” receptors (Fig. 1.1).

1.3 GPCR signalingThe role of GPCRs is to transduce signals from the extracellular to the intra-cellular environment. They are activated by chemically diverse endogenous

11

Figure 1.1. Schematic representation of GPCR topology. GPCRs are characterizedby an N-terminal (N-term) segment, seven transmembrane spanning helices (TM1-7)connected by three extracellular loops (ECLs) and three intracellular loops (ICLs), anda C-terminal (C-term) segment. GPCRs have two main binding sites (BSs): the extra-cellular part of the receptor exposes a ligand binding pocket that recognizes chemicalmessengers and the intracellular part defines the BS for intracellular partners such asheterotrimeric G proteins and β-arrestins.

compounds, e.g. neurotransmitters, biogenic amines, nucleosides, lipids, hor-mones, peptides and proteins, ions, light, and also exogenous compounds [6].When the endogenous agonist binds to the orthosteric (in the extracellular part)binding site (BS), the GPCR undergoes conformational changes in the intra-cellular domain, which result in the activation of signaling pathways.

1.3.1 The canonical G protein pathwayStudies of GPCR-mediated signaling have traditionally focused on the cou-pling to different heterotrimeric G proteins (Fig. 1.2). A heterotrimeric Gprotein is an assembly of a Gα subunit containing the active site, which in-teracts with a Gβγ complex (formed by β and γ subunits). When the agonistactivates the receptor and induces conformational changes, a GDP moleculeis exchanged by GTP within the Gα subunit. The low affinity of the GTP-bound Gα subunit for the Gβγ complex causes the dissociation of the het-erotrimeric G protein, releasing the Gβγ complex from the Gα subunit boundto GTP. Upon release, the Gβγ complex and the GTP-bound Gα subunit canthen stimulate their primary intracellular effectors [6]. There are various Gα,Gβ, and Gγ subunit subtypes, and these interact with specific effectors. For in-stance, Gαs and Gαi/o stimulates and inhibits adenylate cyclase, respectively,the Gαq/11 subunits stimulate protein kinases C and phospholipases C, and theGβγ subunits stimulate membrane proteins such as ion channels [7]. Theseinteractions result in downstream changes in second-messenger levels such ascAMP and inositol triphosphate (IP3) [8, 6].

12

1.3.2 Non-canonical pathwaysResearch during the past decades has shed more light on GPCR signaling andshowed that receptors can also interact with a plethora of intracellular part-ners, including G protein-coupled receptor kinases (GRKs) and β-arrestins(Fig 1.2). To maintain biological homeostasis and quench the levels of sec-ond messengers, GPCRs are phosphorylated by GRKs on the ICLs, resultingin an increased affinity for β-arrestins. This non-canonical pathway leads to re-ceptor desensitization via internalization, degradation or recycling. In additionto receptor desensitization, β-arrestins can also regulate trafficking, signalingafter receptor internalization, and downregulation pathways [9, 10]. Althoughoverlooked, GPCR signaling through β-arrestins affects a variety of key cel-lular functions, including chemotaxis, protein translation, gene transcription,and cellular apoptosis [11]. The discovery of non-canonical GPCR signalingpathways paved the way for the field of biased agonism.

Figure 1.2. GPCRs coupled to intracellular partners. Complexes of a G protein boundto the cannabinoid receptor 1 (CBR1, PDB code: 6N4B [12]) and β-arrestin 2 boundto neurotensin receptor 1 (NTSR1, PDB code: 6PWC [13]).

1.4 GPCR pharmacology1.4.1 Orthosteric ligandsGPCRs possess an orthosteric BS where the endogenous agonist binds to ac-tivate the receptor and induces downstream signaling. In the absence of ag-onists, receptors can exhibit an agonist-independent activity (constitutive ac-tivity). GPCR activity can also be modulated by a wide range of xenobioticcompounds. The efficacy of a ligand is then measured relative to the maximal

13

efficacy of the endogenous agonist. Agonists are categorized as full agonistswhen they achieve the same efficacy as the endogenous agonist and as partialagonists if they reach lower maximal efficacy. In addition to agonists, GPCRscan be modulated by neutral antagonists, which prevent agonists from bind-ing and therefore reduce activity to the basal level. Finally, inverse agonistspromote the deactivation of the receptor by shifting the equilibrium towardsinactive states [14].

The traditional classification of ligands assumes that GPCRs only exist intwo different states: the active and inactive conformation. However, it hasbeen shown that the two-state model cannot capture the complexity of GPCRsignaling. Instead, GPCRs exist in a large spectrum of active and inactiveconformations [15, 16]. Functionally selective ligands have the ability to se-lectively stabilize a subset of these conformations, leading to the coupling tospecific intracellular signaling partners and pathways [17, 16]. The capacityof a ligand to act specifically on a signaling pathway has several therapeuticimplications. For example, morphine, an agonist of opioid receptors, has beenassociated with analgesic effects, but also several side effects including addic-tion and respiratory stress. This can be explained by the fact that morphine is abalanced agonist. Whereas the activation of the G protein pathway is responsi-ble for the analgesic effect, activation of the β-arrestin pathwaymediates manyof the adverse effects. An alternative to morphine is TRV130, a G protein bi-ased ligand that exhibits better analgesic effects and reduced side effects [18](Fig. 1.3). Selective modulation of functional pathways is expected to providedrugs with improved efficacy and safety profiles [19].

1.4.2 Allosteric modulation and GPCR oligomerizationGPCRs can be modulated from sites that are topographically and functionallydistinct from the orthosteric BS [20]. The binding of allosteric modulators withintrinsic activity, e.g. allosteric agonists, triggers conformational changes thatresult in receptor activity modulation even in the absence of an orthosteric lig-and. Positive and negative allosteric modulators (PAMs and NAMs) are ableto affect GPCR signaling only in presence of an orthosteric agonist. Thesemodulators alter the GPCR agonist-bound conformation, resulting in changesin the binding affinity of the agonist for the orthosteric site and/or its efficacy[20, 21, 22]. Several allosteric modulators have been characterized for GPCRs,including small molecules [20, 7] and proteins. G proteins are an example ofprotein allosteric modulators of ligand binding to GPCRs [21]. Moreover, aGPCR can also play the role of allosteric modulators of other GPCRs [23].

Receptor-receptor interactions were first characterized for two class CGPCRs: the γ-aminobutyric acid B receptor 1 and 2 (GABABR1 and

14

Figure 1.3. Schematic representation of functionally selective ligands influence GPCRsignaling. Balanced agonist such as morphine (in orange) can activate both G proteinand β-arrestin pathways. Biased ligands (green and blue) selectively activate path-ways.

GABABR2). The constitutive heterodimerization of these two receptors wasnecessary for the effective signaling and trafficking to the cell surface [24, 25].The constitutive dimerization among class CGPCRs has thereafter opened newperspectives for the study of GPCR pharmacology. Although GPCR pharma-cology has generally assumed that GPCRs work as monomers, a large numberof studies has shown that many receptors form functional dimers or higher-order oligomers in the membrane [26, 27, 28, 29]. Within the oligomers,the protomers can exert positive and negative allosteric modulation on the in-teracting partner [23]. However, understanding of GPCR quaternary struc-tures is still limited [30]. Questions regarding the oligomeric interfaces, theinfluence of oligomers on signaling, and how these can be modulated bysmall molecules are driving current experimental and computational studieson GPCR oligomers. Understanding of the structure and function of GPCRoligomers could contribute to the design of more selective and tissue-specificdrugs [23].

1.5 GPCR structural biologyDetermination of atomic resolution GPCR structures has been challenging.However, the recent structural breakthroughs have led to the determinationof ∼350 GPCR structures, including 64 unique receptors [31, 32].

15

Figure 1.4. Evolution of the number of unique GPCR structures. Figure obtained fromGPCRdb [31].

The two first GPCR structures, Rhodopsin [33] and the β2 adrenergic re-ceptor [34], were determined seven years apart. Difficulties encountered incrystallization were related to the fact that GPCRs are inherently flexible andcan have low native expression levels. Since they are membrane proteins, theirisolation from the membrane in a soluble, functional, and stable formwas chal-lenging [35].

Over the past years, major advances have been made to finally crystallizeGPCRs. To increase receptor solubility and stability outside the membrane,different approaches such as the fusion to soluble proteins, nanobodies, andthermostabilizing mutations have been used. The replacement of the disor-dered ICL3 by fusion proteins (e.g. T4 lysozyme) or nanobodies increasedthe polar surface area of the receptor and reduced receptor flexibility [36, 34].Increased receptor stability was also achieved by thermostabilization. Thisapproach consists in the introduction of mutations in the receptor and the eval-uation of the ability of mutant receptors to bind ligands at increasing tem-peratures. The reduced conformational flexibility of themostable receptors incomparison to the wild-type facilitates the crystallization [37, 38]. This tech-nology, as well as the use of selective nanobodies, has the ability to capture

16

receptors in a specific conformational state, e.g. active or inactive-like states[39, 40]. Moreover, usage of nanobodies allows the study of receptor function,e.g. activation and biased signaling. In addition to these tools, improvementsof crystallization conditions were made [35]. The first β2 adrenergic receptorstructure was crystallized using bicelles [34]. Lipidic cubic phase, where thereceptor is embedded in a membrane-like environment, can also be used inGPCR crystallization [36].

In recent years, cryogenic electron microscopy (cryo-EM) has been appliedto GPCRs [41]. Previously, cryo-EMwas limited to large proteins and resultedin low-resolution structures. With the development of better microscopes, de-tectors, and image processing softwares, GPCR structures can now be deter-mined at atomic resolution [41, 42]. Since the release of the first GPCR cryo-EM structure in 2017 [41], 26 structures from class A, B and F GPCRs havebeen solved.

1.6 New paradigms in GPCR drug discoveryGPCRs play important roles in human physiology. These receptors are in-volved in several diseases, e.g. CNS disorders, cardiovascular diseases, can-cer, asthma, and diabetes [43]. Currently, 34% of FDA-approved drugs me-diate their effects via the modulation of GPCR function [44]. However, tradi-tional drug discovery is failing in delivering effective new therapeutics. More-over, due to the complex nature of diseases, current therapeutic agents exhibitadverse effects and lack of efficacy. With the recent breakthroughs in structuralbiology and the deeper understanding of GPCR structure and function, newparadigms in GPCR pharmacology have emerged [45]. The wealth of GPCRstructures has promoted the use of structure-based methods, e.g. structure-based virtual screening (SBVS) and molecular dynamics (MD) simulations, tostudy activation and identify ligands. The design of novel drugs will requirean integration of the advances in structural biology and pharmacology, e.g.improved understanding of functionally selectivity, polypharmacology, sub-type selectivity, allosteric modulation, and oligomerization. Compounds withtailored properties could lead to more efficacious drugs with less side effects[45, 44, 46, 47, 19, 48].

The work in this thesis will focus on three major areas:• The impact of GPCR structural biology on structure prediction and vir-tual screening (Paper I).

• The application of structure-based methods to design ligands with tai-lored pharmacological properties (Papers II and III).

• Modeling of GPCR heterodimer interfaces ( Paper IV).

17

2. Computational structure prediction ofGPCRs

The past decades have seen an explosion of genome sequencing projects, e.g.the Human Genome Project [49, 50]. This resulted in the sequencing of thenearly complete human genome, yielding approximately 22 333 protein se-quences [51]. As of August 2019, the number of human protein sequencesavailable in the UniProtKB database [52] is around 20 350. The functionof these proteins can often be determined using sequence analysis tools andhomology. However, deeper understanding of their function and interactionswith small molecules will require three-dimensional (3D) information. Exper-imental 3D structures can be obtained by X-ray crystallography, nuclear mag-netic resonance (NMR) or cryo-EM, but the throughput of these techniques islow. As of October 2019, only 6752 unique experimentally determined pro-tein structures for the human species are available in the Protein Data Bank(PDB). To close the gap between the number of available protein sequencesand the number of proteins with 3D structures, computational structure pre-diction methods can be used. Experimental structures are available for only8% of the GPCR family. However, structure-based drug design requires ac-curate models of the receptors [53]. Historically, various structure predictionapproaches have been used to predict of GPCR structure, including ab initiomethods, threading, and homology modeling (HM) [54, 55, 56].

2.1 Ab initioAb initiomethods use protein amino acid sequence to predict protein fold [57],secondary and tertiary structures [58, 59]. The limited structural informationavailable to predict GPCR structure by homology has led to the developmentof ab initio methods for structure prediction [60]. In these methods, the sevenTMs of the receptor are first predicted (using e.g. hydropathicity analysis andmultiple sequence alignment (MSA) profiles) and assembled into a 7TM bun-dle. Energy minimization steps and/or MD simulations are then carried out tofold and refine the protein structure [61, 60, 62].

With the popularity of machine learning applied to life sciences, ab initioand deep learning approaches have been combined to predict protein structures[63]. The AlphaFold system, implemented by the Google DeepMind team, has

18

had great outcomes in the 13th Critical Assessment of Techniques for ProteinStructure Prediction (CASP13). The method ranked first in the free modelinground after the successful prediction of several targets. The Alphafold algo-rithm is based on co-evolutionary data analysis. More specifically, contactmaps of co-evolving residues extracted from MSA of millions of sequenceswere used to fold proteins [64, 65]. Two residual convolutional neural net-works were used. The first one was built and trained on PDB structures topredict interatomic Cβ distances and angles. The second one was trained toscore the geometry and structural accuracy. Protein family-specific potentialsgenerated through pairwise distances were then minimized by gradient descentto predict the protein structure [66, 67].

2.2 ThreadingThreading is based on the observation that the number of protein folds in na-ture is limited [68, 69]. Hence, the sequence in hand is likely to fold as aknown structure regardless of sequence similarity. On these grounds, the goalof the threading algorithm is to perform fold recognition by querying the aminoacid sequence against a library of folds (e.g. CATH or SCOP) and applyinga scoring function to select potential fold candidates for the queried protein[70]. Similar to ab initio methods, co-evolutionary data can be used to guidethreading. Although the global fold might be well predicted using threading,the accuracy of the model might not be sufficient to study protein function anddesign ligands [53].

2.3 Homology modelingHomology modelling (HM) is founded on evolutionary principles, which statethat structure is more conserved than sequence [71]. Hence, proteins sharingsimilar sequences are likely to adopt a similar fold and have a similar function.Using a template structure, the structure of a target sequence can be predictedbased on their pairwise sequence alignment. There are several steps in the pro-cess of protein HM: e.g. template selection, sequence alignment, generation ofprotein core, addition of non-conserved features, build full protein, minimiza-tion, and evaluation of the final models (Fig. 2.1) [72].

2.3.1 Template selection and sequence alignmentThe similarity threshold for two proteins to share the same fold is debated.The empirical rules set the sequence identity cutoff to between 28 and 35%,

19

but this cutoff varies with the sequence alignment length [73]. When the se-quence identity between two sequences (with at least 60 residues) fall underthe 28% threshold, fold conservation is uncertain [73]. There are exceptionsto these rules. In larger protein families such as pentameric ligand-gated ionchannels and GPCRs, the fold in the family is conserved across the differentkingdoms of life despite low sequence identity. Once the global fold of the pro-tein is obtained by homology, the focus is shifted towards local sequence con-servation. In protein families, the members share several motifs and signatures[74, 5]. Two proteins with similar functions should display a conservation ofthe residues with functional roles such as ligand binding, protein recognition,and catalysis. In the GPCR superfamily, MSA of the family has highlightedsequence motifs that are well conserved in the different subfamilies. For in-stance, class A GPCRs have well conserved residues in each TM helix whichare used in different numbering schemes (Ballesteros-Weinstein [75], GenericGPCR numbering [76]) to find relative positions of amino acids across multi-ple sequences. In the Ballesteros-Weinstein [75] numbering scheme, the mostconserved residue of each TM is assigned the number X.50, where X denotesthe TM number. These conserved X.50 positions often have important func-tional roles. For example, the highly conserved Arg3.50 in the DRY motif isinvolved in the ionic lock and the shift of the conformational equilibrium to-wards the inactive state. The Pro7.50 in the NPxxY motif is also predicted toplay important roles in receptor activation [77]. Therefore, both the local andglobal similarities must be considered in template selection (step 1 and 2, Fig.2.1).

2.3.2 Build conserved and non-conserved regionsThe conserved regions correspond to well aligned residues in the alignmentbetween template and target. An initial model of the target, the structural core,is built by copying the coordinates of matched atoms or residues of the tem-plate. Non-conserved regions, e.g. insertions, are modeled as loops (step 5,Fig. 2.1). Finally, energy minimization (step 6, Fig. 2.1) is applied to correctgeometry errors and refine the overall model [72].

2.3.3 Model evaluationGenerated models can be evaluated for their geometry and stereochemistryusing knowledge-based scoring functions (step 7, Fig. 2.1). Programs suchas PROCHECK [78] compare the geometry of models to that of high-qualityexperimental structures. The discrete optimized protein energy (DOPE) [79],an atomic distance-dependent statistical potential derived from 3D structuresdatabases, can be used to assess model quality. In addition to structural ac-curacy, models can be evaluated for their ability to recognize known ligands.

20

This allows the selection of reliable BS conformations for ligand discoverypurposes [80, 81].

Start of HM

1. Search for templates

2. Sequence alignment

3. Extraction of homology fea-tures, e.g. backbone, structurallyconserved regions, spatial restraints

4. Build structural coreand add side chains

Core good:Yes or No ?

5. Add loops for insertion/deletionsand explore rotamer libraryfor non-conserved residues

6. Optimize model: local and globalminimization using MD, simulatedannealing or conjugated gradient

7. Evaluate model

Model good:Yes or No ?

End of HM

Edit the alignment

Edit the core

start

Yes

Yes

No

No

Figure 2.1. Summary of the key steps of the HM process.

21

2.3.4 Homology modeling with MODELLERIn this thesis, HM of proteins was performed using MODELLER [82]. MOD-ELLER builds homology models by satisfying spatial restraints [82]. The spa-tial restraints are encoded as probability density function built upon geomet-rical features, e.g. interatomic distances and torsions, derived from the align-ment between the target and the template structure. The probability densityfunction also includes stereochemical restraints (e.g. bond lengths, angles, tor-sion angle preferences, and non-bonded interaction distances) extracted frommolecular mechanics force fields and a database of protein structures. Gen-erated models are then minimized using conjugated gradient and MD simu-lations are used to reduce violations of spatial restraints. Models quality isfinally evaluated by scoring functions, e.g. DOPE score [79].

2.4 Protein-protein dockingProtein-protein interactions play key roles in biological functions such as sig-naling and transport. The elucidation of receptor-receptor interfaces and struc-tural details of the interactions could guide the design of oligomer-specificligands [83, 84]. Oligomeric complex formed by multiple protein monomerscan be predicted using protein-protein docking. The docking process consistsof two steps: the sampling and the scoring.

2.4.1 The samplingDuring the sampling step, possible binding modes for the two monomers areexplored using rigid-body docking or conformational sampling. Prediction ofthe complex can be made using different algorithms: exhaustive global search,local shape feature-matching, or randomized search. In the exhaustive globalsearch, a protein with three rotational and translational degrees of freedom ismoved around a static protein to sample potential complexes. The extensivesearch is made possible by use of grid-based approaches such as fast Fouriertransformations or accelerated search in a cartesian grid space. In the localshape feature-matching approach, local complementarity between molecularsurfaces of the docking partners is sought. Finally, in the randomized search, amonomer is first randomly placed in the vicinity of a fixed monomer BS. Then,binding orientations are sampled using Monte Carlo simulations or geneticalgorithms [85].

2.4.2 The scoringGenerally, the sampling and scoring steps are decoupled. A large number ofcomplexes are first generated and then the scoring step is carried out. Docked

22

complexes are scored using empirical-based, knowledge-based, or force field-based scoring functions. Docking algorithms and their scoring functions areregularly evaluated for their performances on benchmark sets [85] during Crit-ical Assessment of Predicted Interactions (CAPRI) competitions [86]. Thesebenchmarking sets contain both unbound and bound experimental structuresof protein-protein complexes in order to evaluate the ability of the dockingalgorithm to account for conformational changes upon binding.

2.4.3 Protein-protein docking with HADDOCKThe protein-protein docking program HADDOCK [87] has been extensivelyevaluated in the CAPRI competitions [88, 86, 89, 90, 91, 92]. HADDOCKusesa randomized search algorithm and requires a set of ambiguous interaction re-straints (AIR) to guide the docking of the two monomers. A sequence of rigid-body minimization, binding conformation optimization with flexible atoms atthe interface, and a short MD simulation is performed. Finally, clustering isused to provide the best docking candidates [87]. In this thesis, protein-proteininterfaces were predicted using homology-based information, i.e. the 3D struc-ture of homologous interfaces, andmutagenesis data. The docking calculationswere performed using HADDOCK and the best docking candidates were re-fined using MD simulations.

23

3. Molecular dynamics simulation

Protein structures are dynamic. The interactions between a protein and ligandinvolve induced-fit mechanisms [93] and/or the selection of receptor confor-mations by ligands [94]. GPCRs have a plethora of active and inactive confor-mations which can be stabilized by functionally selective ligands [95]. There-fore, in order to understand protein function, flexibility must be considered.However, the majority of structure prediction methods described in previouschapters provide static protein structures. Experimental structure determina-tion approaches such as NMR [96] or time-resolved crystallography [97] canbe used to explore the motion of proteins. Besides, biomolecular motions canbe modeled computationally using MD simulations [98, 99]. In MD simula-tions, interactions between particles are derived using molecular mechanicsforce fields and the time-dependent trajectory of particles can be obtained bynumerical integration of Newton’s equations of motion.

3.1 The force fieldMolecular mechanics force fields are used to define the potential energy U ofa system of N interacting particles, e.g. protein atoms. Parametrization offorce fields is generally performed by quantum mechanics (QM) calculationsor fitting to experimental data. Several force fields for biomolecular simula-tions, which differ mainly in the parametrization scheme are available (e.g.CHARMM [100], OPLS-AA [101] or AMBER [102]). A typical force fieldis derived by the summation of bonded (bonds, angles, improper torsions, anddihedrals, see Eq. 3.2) and non-bonded (electrostatic and van der Waals inter-actions, see Eq. 3.3) contributions to the potential energy.

U = Ubonded + Unon−bonded (3.1)

3.1.1 The bonded interactionsThe intramolecular interactions between a set of bonded particles can be de-scribed by harmonic potentials and periodic functions. The contributions to the

24

potential energy from bond-stretching, angle-bending, improper torsion angleand dihedral angle deviation can, for example, be derived from Eq. 3.2:

Ubonded = Ubonds + Uangles + Utorsions + Udihedrals + Uimpropers

=∑bonds

kAB

2(lAB − lAB,e)

2 +∑

angles

kABC

2(θABC − θABC,e)

2

+∑

dihedrals

kϕ2(1 + (cos(nϕ− γ))) +

∑impropers

kζ2(ζ − ζe)

2

(3.2)

where kAB , kABC and kζ are the force constants applied to bond stretching,angle bending and improper torsion in the harmonic potential when they devi-ate from their reference length (lAB,e), angle (θABC,e), and improper torsionangle (ζe), respectively. For the periodic function of the dihedral term, kϕ isthe energy barrier, γ the phase shift, and n the periodicity of the dihedral angleϕ.

3.1.2 The non-bonded interactionsThe non-bonded terms describe the intermolecular interactions between par-ticles of the system. The contributions to the potential energy accountingfor these interactions are the electrostatic and the van der Waals terms. TheCoulomb potential (Uelec) describes the electrostatic interactions between twoparticles i and j with charges qi and qj , respectively, as a function of thedistance lij . The van der Waals term describes the interactions betweennon-bonded particles, which when close in space become transient dipoles(induced-dipoles). Generally, a Lennard-Jones potential (UL−J ) is used tomodel these interactions using an attractive component (exp − 6 potential)for London dispersion-attraction forces and a repulsive component (exp− 12potential) to describe repulsive forces due to overlapping orbitals and the Pauliexclusion principle. The potential energy of the system for non-bonded inter-actions Unon−bonded can therefore be defined as:

Unon−bonded = Uelec + UL−J

=∑

charges

qiqj4πϵ0lij

+∑L−J

4ϵij

((σijlij

)12

−(σijlij

)6)

(3.3)

where qi and qj are the charges of particles i and j, ϵ0 is the vacuum orreference medium permittivity, ϵij the depth of the well, σij the distance atwhich the potential is equal to 0 and lij the distance between the particles i andj.

25

Figure 3.1. Bonded and non-bonded interactions in an arginine residue.

3.2 The simulation3.2.1 Initial parameters and settingsAs we will see later, MD simulations consist in integration Newton’s equationsof motion at each time-step δt. To be able to capture all motions of the system,including high frequency motions such as bond vibrations, a small δt (e.g. 1-2 fs) is set. To start the simulation, initial velocities (e.g. extracted from aMaxwell Boltzmann distribution) are assigned to all particles of the solvatedsystem. The simulation can be performed with different boundary conditions.In this thesis, we simulated our solvated protein systems using periodic bound-ary conditions (PBC). To account for the fact that particles near the edges ofthe simulation box experience different forces from those near the center, alarge number of identical copies of the simulation box is generated in all di-rections. This continuum allows particles at the edge of the box to interactwith particles in the neighboring box. When particles exit a simulation box,an image enters from the opposite edge keeping the total number of particlesconstant. A few precautions are necessary when using PBC. The box must belarge enough to prevent a particle from interacting with its periodic image andcutoff distances must be applied for long-range interactions. For instance, inACEMD [103], cutoff distances complemented by a smooth switching func-tion are applied for Lennard-Jones and short-range electrostatic interactions,whereas long-range electrostatic interactions are calculated using the particle-mesh Ewald algorithm [104]. A similar protocol can be used in GROMACS[105].

26

3.2.2 Equations of motionFor each time-step δt, a new configuration of the system is generated. Bycalculating the gradient of U (∇U ), the force applied on each particle i can bederived using Newton’s law:

F⃗i = −∇iU

= mid2r⃗idt2

= midv⃗idt

= mia⃗i

(3.4)

Once the force F⃗i is obtained using Eq. 3.4, Newton’s equation of motioncan be solved using Taylor expansion with different integration schemes. Forinstance, the classical Verlet algorithm [106] uses the acceleration ai(t) at timet and the positions ri at time t and t− δt to derive ri(t+ δt) (Eq. 3.5):

ri(t+ δt) = 2ri(t)− ri(t− δt) + ai(t)δt2 (3.5)

Subsequently, the velocity vi(t) can be calculated with the following equation:

vi(t) =ri(t+ δt)− ri(t− δt)

2δt(3.6)

The leap-frog algorithm uses the half-step velocity vi(t +12δt) obtained

using the acceleration ai(t) (Eq. 3.7) and the current position ri(t) to obtainri(t+ δt) (Eq. 3.8):

vi(t+1

2δt) = vi(t−

1

2δt) + ai(t)δt (3.7)

ri(t+ δt) = ri(t) + vi(t+1

2δt)δt (3.8)

After each time-step δt, the force F⃗i is derived from the potential energy ofthe system (Eq. 3.4). Positions are saved in trajectory files and a new iterationof the algorithm is initiated.

27

4. Structure-based virtual screening

The increasing availability of protein structures and computational resourcesled to the rise of structure-based drug design [107, 108]. Traditionally, largelibraries of candidate molecules have been screened in high-throughput screen-ing (HTS) campaigns. However, structure-based virtual screening (SBVS)methods are now routinely used in drug discovery campaigns as the startingapproach to identify lead candidates in large libraries of compounds [108].Unlike experimental HTS, this approach is fast and cost efficient. Millions ofmolecules can rapidly be screened at supercomputer clusters and the hit ratesare often greater than from HTS [109].

4.1 Molecular dockingThe aim of molecular docking is to predict the binding mode of a ligand in aprotein BS. Thus, the 3D coordinates of the protein, obtained either by com-putational modeling or from experimental methods, are required as well asknowledge about BS location. Docking of a ligand can be divided into twosteps: posing (or sampling) and scoring.

4.1.1 Receptor binding site preparationPrior to docking calculations, the receptor BS must be prepared. The proto-nation states of ionizable residues are assigned with careful attention on BSresidues. Moreover, functional water molecules [110] and co-factors presentin the BS can be kept in the docking calculations as in Paper II.

4.1.2 Library preparationIn prospective virtual screening (VS), the sets of compounds to be dockedare often extracted from commercial chemical libraries with millions of com-pounds such as the ZINC database [111]. Prior to screening, libraries arefiltered based on physicochemical properties, e.g. molecular weight (MW),lipophilicity (logP), and the number of hydrogen bond acceptors and donors.Moreover, known pan-assay interfering compounds (PAINS) [112] are filteredout. With the increasing need for ligands with tailored properties, focused li-braries are generated by extracting only a subset of compounds from large

28

databases. For instance, in Paper III, we extracted a set of fragments to beattached to an amidopropyl or oxobutyl linker of the 2,3-dichlorophenyl piper-azinemoiety, a scaffold recognized by dopamine receptors [113]. The obtainedfocused library contained compounds with high probability of binding to theD2 dopamine receptor (D2R). Finally, library generation involves considera-tion of relevant protonation and tautomeric states, e.g. at physiological pH, forall compounds.

4.1.3 The molecular docking process: sampling and scoringThe sampling, or posing, step aims at generating a large number of orientationsand conformations of the ligand in the BS. These conformations can be sam-pled using different algorithms, e.g. genetic algorithms, Monte Carlo simula-tions, or anchor-and-grow approaches [114]. For each sampled conformation,a scoring function is used to estimate the binding energy or the fitness of abinding mode. Empirical scoring functions assess how well binding energiesreplicate experimental data. The free energy is estimated by summation of thecontributions from uncorrelated terms such as hydrogen bonding, hydropho-bic interactions, ligand entropy, and lipophilic interactions [115]. Knowledge-based scoring functions assess how well binding modes and conformationsreplicate experimental complexes. With these functions, protein-ligand com-plexes are modeled using simplified pairwise potentials such as the potentialmean force [116]. Finally, force field-based scoring functions predict the bind-ing affinity of ligands for the receptor as an energy score. This energy is de-rived from the summation of molecular mechanics force fields terms such asvan der Waals and electrostatic interactions (see Eq. 4.1) [117].

4.1.4 Retrospective screeningMolecular docking can be used to perform retrospective screening to evaluatethe performance of the software, e.g. by redocking a co-crystallized ligandto its BS. Moreover, it can be employed to assess the accuracy of predictedstructures by measuring the ability of models to recognize known ligands. Toperform such retrospective docking screens, a set of known ligands with a bi-ological activity for the target protein must be compiled. These ligands areoften extracted from databases with bioactivity data such as ChEMBL [118].In addition to active ligands, a set of non-active compounds is used as decoys.Decoys can be identified experimentally, e.g. compounds that were inactivein HTS, or generated computationally, e.g. using property-matched decoys[119]. The sets of actives and decoys are screened against the structure, andthen the ranking results are assessed. Ligand enrichment describes how wellthe protein model identifies active compounds among decoys. In this thesis,we used the area under the ROC curve (AUC), which represents the percentage

29

of found actives (y-axis) against the percentage of decoys screened (x-axis).Integration of the AUC allows to quantify ligand enrichment by the receptorstructure. For a model to be used in prospective screening, early recognitionof actives is advantageous. To favor early enrichment, a logarithmic transfor-mation of the AUC [120] can be applied. In this work, we used the logarithmof the AUC to measure ligand enrichment (Fig. 4.1).

Figure 4.1. Example of ligand enrichment curve. ROC curve for a database of ligandsand property-matched decoys ranked bymolecular docking. The percentage of ligandsand decoys identified are shown on the y- and x-axis, respectively. The solid black linerepresents random enrichment of ligands.

4.1.5 Molecular docking with DOCKMolecular docking calculations in this thesis were performed with DOCK[117]. In DOCK, the protein is held rigid and the BS must be predefined, oftenusing a co-crystallized ligand. The docking process consists of an anchor-and-grow approach to sample ligand conformations and a force field-based scoringfunction to evaluate binding modes. During the anchor-and-grow process, therigid core of the ligand is docked in the BS, followed by a growing step whereflexible branches of the ligand are added and sampled in the BS. Docking posesare then evaluated based on electrostatic and van derWaals interactions as wellas their desolvation penalties:

Ebind = Evdw + Eelec − Edesolv (4.1)

30

where Evdw and Eelec are the van der Waals and electrostatic energy contri-butions to the binding energy Ebind, and Edesolv is the desolvation penalty oftransferring the ligand from bulk solvent to the BS with dielectric constantsϵsolvent = 78 and ϵprotein = 2, respectively.

When the full library of conformers has been docked and minimized, theseare scored and ranked. The conformation with the lowest energy is retrievedas the predicted binding mode of the ligand.

31

5. Paper I: Performance of Virtual Screeningagainst GPCR Homology Models: Impactof Template Selection and Treatment ofBinding Site Plasticity

Despite tremendous breakthroughs in crystallization and cryo-EM [121, 35,32], 3D structures are still missing for more than 80% of GPCR family. Thisshortage of atomic resolution structures has led researchers to rely on compu-tationally predicted models for their structure-based drug discovery campaigns[108, 80, 81, 109]. While HM was the most favored and successful structureprediction method in GPCR Dock competitions [54, 55, 56], no guidelinesemerged for the use of HM for VS purposes. Various studies have attemptedto benchmark homologymodels for their suitability in VS [122, 123, 124, 125].However, questions such as how to select templates and how to use homologymodels in VS campaigns remained unanswered.

In Paper I, we investigated the impact of template selection on the VS per-formance of homology models of the D2R and serotonin 5-HT2A receptors (5-HT2AR). Both receptors belong to the aminergic subfamily of class A GPCRsand are involved in many CNS disorders, e.g. schizophrenia and Parkinson’sdisease. Homology models of the two targets were generated based on 12aminergic and four non-aminergic GPCR templates. Selected templates cov-ered a large span of TM and BS sequence identity, ranging from 21 to 77% andfrom 6 to 94%, respectively. Based on each template, a set of 250 homologymodels was generated for each target, from which 50 were selected based onDOPE score [79] for analysis.

The evaluation of templates using the average root mean square deviation(RMSD) of homology models to the crystal structures [126, 127] suggestedan improved structural accuracy of TM and BS regions when comparing tem-plates with low (<30%) and high (>50%) sequence identity to the target. How-ever, templates with sequence identity between 30-50% resulted in modelswith variable accuracy.

To assess template performance in VS, molecular docking screens of knownligands and generated decoys were carried out against homology models. Foreach template, we extracted the distribution of ligand enrichment yielded by

32

the set of 50 homology models. There was a large variation in ligand enrich-ment for models based on the same template, reflecting that different BS con-formations were explored by the homology models. This led us to consider themedian ligand enrichment as a better descriptor of template performance (Fig.5.1). Moreover, we considered BS flexibility by deriving a ligand enrichmentfor an ensemble of homology models. The ensemble enrichments were oftengreater than the median enrichments, but did not outperform the best enrichingmodels.

Figure 5.1. Ligand enrichment by homology models based on 16 templates. Distri-bution of aLogAUC values for (A) D2R and (B) 5-HT2AR models based on differenttemplates. Distributions are shown using a boxplot representation. Each boxplot de-scribes the results for 50 models obtained based on one template. The last boxplot ingray corresponds to the results of all 600 models based on different aminergic receptortemplates. Ensemble enrichments are represented by red filled circles.

Our work identified several weaknesses of the use of homology models inVS. Some distant templates were able to provide models with good median en-richment. However, we observed that the binding modes of top-ranked com-pounds did not capture the charge-charge interaction with D3.32 and extendtowards TM5 with an aromatic moiety as typically seen for aminergic ligands.Moreover, evaluation of template bias highlighted that models preferably en-riched ligands similar to the co-crystallized ligand of the template. The chemo-type bias was reduced by considering an ensemble of structures. A strongerenrichment of all chemotypes and full set of ligands were observed for the en-semble of two models based on different templates.

The modeling of ECL2 using HM is challenging, in particular if the numberof templates is small [54, 55, 56, 128]. We analyzed the impact of ECL2on model performance in VS. The median ligand enrichment considerablyimproved if the loop was removed. These improvements were more pro-

33

nounced for D2R, for which the crystal structure showed a loop conformationnot present in the template set. Although decoupling the impact of the loopfrom the TM contributions to ligand enrichment is not trivial, these resultssuggested a more careful modeling of the ECL2 is needed.

The relationships between sequence identity, median ligand enrichment,and structural accuracy of the BS were then analyzed. We observed a moderatenegative correlation between BS accuracy and ligand enrichment, suggestingthat better BS representation leads to better VS performance. In addition, tem-plates with low BS sequence identity yielded lower median ligand enrichments(poor to fair) than templates with higher sequence identity which generallyproduced good to excellent enrichments. However, templates sharing >45%sequence identity with the target tended to have similar ligand enrichment val-ues. These results can be explained by many factors, including the fact thataminergic receptors are characterized by the conserved D3.32, which recog-nizes compounds with a positively charged amine. Despite the differences inBS composition, this conserved residue might drive ligand binding.

Homology models of the two target receptors were finally refined by MDsimulations. As for the homology models, sets of MD snapshots were eval-uated for their structural accuracy and ligand enrichment. The MD-refinedmodels did not lead to improved structural accuracy or better VS performancein comparison to the raw homology models.

Overall, the results of Paper I suggest that the best template is not nec-essarily the closest. Although, templates with higher sequence identity wereexpected to lead to better enriching models, the highest enrichment and mostaccurate BS were not always based on the closest homolog. Nonetheless, themoderate correlation between BS accuracy and ligand enrichment suggestedthat molecular docking could be used to identify good representations of theBS. While incorporation of BS flexibility using ensembles of models did notimprove VS performance, it often resulted in enrichment greater than the me-dian. In addition, using an ensemble of models based on different templatesallowed to reduce template bias. Finally, the refinement of homology modelsby MD simulations did not improve VS performances.

34

6. Paper II: Docking Screens for DualInhibitors of Disparate Drug Targets forParkinson’s Disease

Parkinson’s disease (PD) is a neurodegenerative disorder characterized by aprogressive loss of dopaminergic neurons in the substantia nigra pars com-pacta. Patients suffering from PD display motor symptoms such as stiffness,bradkynesia and tremor, but also non-motor symptoms such as cognitive de-cline and depression [129]. They are traditionally treated with the dopamineprecursor Levodopa. However, the loss of efficacy of Levodopa and the manyside effects incurred after long term use [130] have led to a need for novel ther-apeutics [131]. Because PD is a multifactorial disease, a polypharmacology-based approach has been proposed for the design of new therapeutics [47,132]. Polypharmacology has gained popularity for the treatment of CNS dis-orders after the discovery that Clozapine superiority in clinical treatment ofschizophrenia was due to its complex pharmacological profile [46]. As a re-sult, numerous antipsychotic drugs with such complex profiles have been ap-proved by the FDA since the late 2000s [133].

Studies on PD have suggested various potential non-dopaminergic targetcandidates, including enzymes and other GPCRs [131, 134]. The A2A adeno-sine receptor (A2AAR), a nucleoside GPCR, and the monoamine oxidase B(MAO-B), an enzyme, are two targets proposed as a potential dual-target can-didates after the serendipitous discovery that caffeine derivatives (e.g. CSC),which are antagonists of A2AAR, could also inhibit MAO-B [135]. A dualinhibitor of A2AAR and MAO-B would combine the neuroprotective effectsof A2AAR antagonism [136] with a sustained dopaminergic signaling throughinhibition of MAO-B [137].

In Paper II, we investigated if SBVS could be used to identify dual-targetA2AAR /MAO-B ligands. We first screened two chemical libraries containing0.8 million fragment-like and 4.6 million lead-like compounds from the ZINCdatabase [111] against both A2AAR [38] and MAO-B [138] crystal structures.A set of 11 lead-like and 13 fragment-like compounds was then extracted fromthe top 500 compounds with the lowest consensus scores.

Interestingly, although the A2AAR and MAO-B binding sites were distinctand no significant similarity was found between known A2AAR and MAO-B

35

Figure 6.1. Summary of molecular docking results for the identification dual-targetcompounds. We experimentally evaluated the 24 predicted dual-target compounds atboth A2AAR andMAO-B. Six compounds displaced the radioligand (>60%) at 30 µM(hit rate of 25%) in A2AAR binding assays (Ki values ranging from 19 to 7100 nM).Twelve compounds showed >70% inhibition at 30 µM (hit rate of 50%) in MAO-Benzymatic assays (IC50 values ranging from 61 to 8700 nM). Among the confirmedhits, four compounds elicited activity at both A2AAR and MAO-B (hit rate of 17%).The most potent dual-target ligand had a Ki of 19 nM at A2AAR and an IC50 of 100nM for MAO-B.

ligands (Tc < 0.3 for 99.9 % of the compounds), the predicted dual-ligandsexhibited similarities in their shapes and binding modes. Compounds 1 and 3in A2AAR (Fig 6.2.a and c) and MAO-B (Fig 6.2.b and d) were anchored byhydrogen bonds to a central amino acid in both BSs (N2536.55 in A2AAR andY326 in MAO-B). In addition to hydrogen bonding, the dual-target ligandswere also well anchored in both binding pockets by clusters of hydrophobicresidues. To summarize, the same ligand functional groups interact with simi-lar protein residues in the two BS. [139]. Despite the structural dissimilarities,

36

the docking scoring function was able to capture similarities in the BS shapeand polarity.

Figure 6.2. Predicted binding modes of two dual-target ligands (compounds 1 and3) and structure activity relationships (SAR). The binding modes of compound 1 (a,b) and compound 3 (c, d) in the A2AAR (grey cartoon) and MAO-B (green cartoon)BSs. SAR analyses were performed by selecting and evaluating a set of analogs ex-perimentally. The grey and green bars represent the fold changes in affinity and IC50.

Potential concerns in the development of multi-target ligands are the dif-ficulties in optimization and the tendency of screens to identify promiscuouscompounds [140, 133]. Analysis of PubChem bioassay records showed thatcompounds 1 and 3 did not display activity in the majority of HTS campaignsdespite being assayed extensively [141]. Both dual inhibitors were selectivefor A2AAR and MAO-B over other drug targets and are thus not promiscu-ous ligands. To understand the SAR of compounds 1 and 3 at both A2AARand MAO-B, we performed an analog search by catalog. Selected analogswere predicted to preserve the binding modes observed for compounds 1 and

37

3, which was confirmed for compound 3 by a crystal structure of an analogbound to the A2AAR [142]. Out of nine analogs of compound 1, only com-pound 1a displayed an improvement at both A2AAR (27-fold) and MAO-B(12-fold) (Fig 6.2). Similarly, only two analogs of compound 3 maintainedactivity at both targets. Replacement of the ester moiety by an amide groupresulted in loss of activity at MAO-B (Fig 6.2). This result suggested that theester moiety which was predicted to bind in a narrow channel connecting twosubpockets of the MAO-B BS, could not be substituted by a more rigid group(e.g amide). Furthermore, the most potent dual-target ligands, 1a and 3, wereconfirmed as selective A2AAR antagonists and MAO-B inhibitors over othersubtypes.

Finally, we tested the two rationally designed dual-target ligands ondopaminergic neuronal-like SH-SY5Y cells pre-exposed to the 6-OHDA neu-rotoxin for their cytoprotective effects. Both compounds 1a and 3 displayedneuroprotective effects and reduced cell death by ∼14% which is similar tothe effect of the known dual-inhibitor CSC.

While the design of multi-target compounds has often relied on known scaf-folds, we carried out an unbiased VS to explore the availability of dual-activitycompounds in commercial chemical space. The focus on the lead-like andfragment libraries allowed us to find compounds with properties compatiblewith oral availability and blood-brain barrier penetration. Moreover, numer-ous structures were available for both A2AAR and MAO-B, allowing the ex-ploration of different BS conformations for each target. The difficulties in op-timization of dual-target activity suggested that high-level atomic details arerequired to develop lead candidates. In the case of A2AAR and MAO-B, thesame functional groups in the ligand interacted with similar protein residuesin the two BSs. However, when the BS of the two targets are very dissimilar,e.g. in shape and polarity, different functional groups of the ligands will be in-volved [139]. For such scenarios, SBVS needs to be complemented with otherapproaches such as the design of focused libraries (Paper III).

The results of Paper II show that SBVS can be used for the identificationand optimization of novel dual-target ligands. The design of compounds withpolypharmacological profile could provide starting points for development ofimproved drugs against diseases with complex etiology.

38

7. Paper III: Structure-Guided Screening forFunctionally Selective D2 DopamineReceptor Ligands from a Virtual ChemicalLibrary

GPCRs exist in a plethora of active conformations which can be stabilized bydifferent ligands. Although themolecular mechanism of biased signaling is notfully understood, functionally selective ligands (or biased ligands) can specif-ically modulate the canonical G protein-dependent or non-canonical signalingpathways by stabilizing distinct receptor conformations [143, 95, 144]. Theuse of morphine, an opioid ligand known for its analgesics effects, has beenlinked to several side effects including dependence and respiratory stress. Thisduality in morphine action is due to its functional promiscuity. Morphine isable to activate both the G protein pathway, which is responsible of the anal-gesic effect, and the β-arrestin pathway which mediates many of the adverseeffects. The design of functionally selective ligands, e.g. TRV130 [18], is ex-pected to provide safer and more effective drugs [19].

In Paper III, we investigated if SBVS could be used to identify function-ally selective D2R ligands from a tailored library. A D2R homology modelwas selected based on its enrichment of known biased ligands. During this ret-rospective screening, a secondary binding pocket where selective compoundsextendedwas identified. In this secondary site formed by TM1, TM2 and TM7,polar contacts between biased ligands and the Glu952.65 and Ser4097.36 werefrequently observed.

Because of the molecular complexity of bitopic compounds, they are likelynot well covered in commercial libraries. Therefore, their identification byHTS or SBVS is improbable. We thus opted for the design of a focused libraryof D2R ligands that potentially bind to both orthosteric and secondary BSs. Thefocused virtual chemical library was built using a 2,3-dichlorophenyl piper-azine moiety, a privileged structure recognized by dopamine receptors [113],as anchor connected to either an oxobutyl or amidopropyl linker. Fragmentswith different functional groups, sizes and shapes, but with carboxylate and hy-droxyl termini were extracted from the ZINC database [111]. These fragmentswere attached to the amidopropyl (for the carboxylate group) or oxobutyl (forthe hydroxyl group) linker using amide coupling and nucleophilic substitu-tion, respectively. After reactions, two libraries which yielded a total of 12

39

985 unique compounds, were docked to the D2R homology model (Fig. 7.1)and the 300 top-ranked compounds of each set were visually inspected. A setof 18 compounds with the piperazine moiety well anchored in the orthostericsite and forming salt-bridge interactions with D1143.32 was prioritized. In ad-dition, these compounds extended into a secondary binding pocket formed byTM1, TM2 and TM7. All selected compounds were successfully synthesizedand tested for their signaling properties.

Figure 7.1. Design of the virtual chemical library of potential D2R biased ligands.The functionally selective ligands were designed using an anchoring moiety in theorthosteric site and a linker (cyan) that allowed extension of a fragment (pink) to asecondary binding pocket.

Canonical G protein activation was assessed in [35S]GTPγS binding assays,whereas β-arrestin 2 recruitment was characterized using the DiscoveRx Path-Hunter assay. All compounds were first tested at a single-point concentra-tion (10 µM) to estimate the Emax. The full agonist (Emax = 100%) Quinpi-role, with a balanced profile, and the β-arrestin 2 recruitment partial agonistUNC0006 (Emax = 18%) were used as references. All 18 compounds actedas partial agonists of at least one signaling pathway (Fig. 7.2). In addition,single-point experiments demonstrated that a majority of the predicted biasedcompounds acted as balanced compounds and did not show clear preferencefor G protein or β-arrestin 2 signaling. For this reason, no relation betweenstructural features of the moiety in the secondary pocket and the functional se-lectivity profile of the partial agonists could be derived.

A set of six compounds with good efficacy or a preference for the β-arrestin2 recruitment assay were selected for full dose-response curve determination.

40

Similar to the reference β-arrestin 2 partial agonist UNC0006, compound 4exhibited an EC50 = 320 nM and Emax = 16 % for β-arrestin 2 recruitment butdid not promote [35S]GTPγS binding. Another five compounds were partialagonists (EC50 from 30 to 600 nM) of both G protein binding (Emax 4.5-24%)and β-arrestin 2 recruitment (Emax 16-40%), but displayed preferences for theβ-arrestin signaling pathway.

Figure 7.2. Comparison between the Emax values of the 18 compounds and UNC0006in β-arrestin 2 recruitment and [35S]GTPγS binding assays. Compounds with oxy-butylene linker and the reference compound, UNC0006, are shown with filled greencircles. The amidopropyl containing compounds are represented with blue circles.

In this proof-of-concept study, we were able to explore a specific region ofchemical space to identify a set of synthetically tractable compounds. How-ever, the lack of SAR still limits the understanding of functional selectivity.Despite that all compounds were selected based on the interactions observedfor known biased ligands, a majority displayed a balanced signaling profile.A potential explanation is that compounds were identified using D2R struc-ture modeled based on the inactive conformation of the D3R. The identifiedinteractions might not be sufficient to stabilize the relevant subset of activeconformations. Further structural information is needed. With the recent re-lease of the D2R structure, MD simulations could be used to generate multiple

41

active state conformations to better understand functional bias at the molecularlevel.

To summarize, in Paper IIIwe explored the design of functionally selectiveD2R ligands using SBVS and a tailored virtual library. This work showcasesthe design of novel ligands by combining computational approaches (SBVSand virtual library design), chemical synthesis, and biological testing. How-ever, rational design of biased ligands is still limited by the lack of structuralinformation.

42

8. Paper IV: Mapping the Interface of a GPCRDimer: A Structural Model of A2A

Adenosine and D2 Dopamine ReceptorHeteromer

Dopaminergic neurons in the substantia nigra pars compacta project in thestriatum, which also selectively express the A2AAR [145]. The existence ofA2AAR-D2R heterodimers in the striatum [146] was supported by the iden-tification of a negative allosteric modulation of D2R by A2AAR activation[147]. The A2AAR-D2R heterodimer thus represents a potential drug targetfor diseases caused by D2R signaling impairment such as PD. Blockade of thenegative allosteric modulation exerted on D2R by A2AAR activation could beprevented using A2AAR antagonist, e.g. caffeine, leading to neuroprotection[136] (see Paper II).

The structural basis of A2AAR-D2R heterodimerization is not well under-stood. In Paper IV, we attempted to characterize the dimer interface usingbiophysical experiments in combination with computational approaches. In aprevious study [148], synthetic TM helices of the D2R (TM-ID2 to TM-VIID2)were employed to competitively inhibit A2AAR-D2R dimerization. WhereasTM-ID2-TM-IIID2 and TM-VIID2 peptides did not affect heterodimerization,TM-IVD2-TM-VID2 induced a concentration-dependent decrease of BRET1

signal. These results suggested the involvement of TM-IVD2 and TM-VD2

in the A2AAR-D2R interface. Similarly, in Paper IV, we investigated the in-volvement of TM helices of the A2AAR in the dimer interface. BRET1 ex-periments were carried out using TM helices of the A2AAR (TM-IA2A to TM-VIIA2A). TM-IVA2A-TM-VIA2A displayed a concentration-dependent inhibi-tion of the BRET1 signal (Fig. 8.1). This result was also confirmed in PLA,where TM-IVA2A and TM-VA2A disrupted A2AAR-D2R dimerization in dif-ferent cell systems.

Finally, the TM helices were assayed for their impact on the allosteric mod-ulation in the A2AAR-D2R heterodimer. The A2AAR agonist CGS-21680 re-duced the affinity of the D2R agonist Quinpirole in membrane preparations ex-pressing A2AAR-D2R heterodimer. However, in the presence of TM-IVA2A

and TM-VA2A, the negative allosteric modulation of A2AAR activation onD2R was significantly reduced.

43

Figure 8.1. Effects of A2AAR TM helices on A2AAR-D2R dimerization. Theconcentration-response effect of the individual A2AAR TM helices on BRET1 signalsof A2AAR-D2R heterodimer.

These experimental results were subsequently used to guide protein-proteindocking and modeling the A2AAR-D2R heterodimer. Various orientations ofthe A2AAR and D2R protomers were sampled and two hundred models of theA2AAR-D2R dimer were generated and clustered based on interface RMSD.One cluster contained dimer conformations involving extensive interactionsbetween TM4 and TM5 of A2AAR and D2R. The predicted A2AAR-D2R in-terface had a total BSA of ∼1,200 Å. The modeled heterodimer structure wasthen refined using MD simulations. Analysis of the MD snapshots highlighteda few important interactions. The residue Y1925.41 in D2R participated instacking interactions with a cluster of aromatic residues in TM5 of A2AAR(Y1795.40 and F1835.44). In the intracellular part of the receptors, L2075.56in D2R interacted with I1274.48, V1305.51, and L1314.52 in A2AAR. Finally,K2115.60 in D2R was in contact with I1274.48 in A2AAR (Fig 8.2).

To evaluate the importance of these interactions for dimerization, mutagene-sis studies were performed for the D2R. The single-mutant Y1925.41A and dou-ble mutant L2075.56A / K2115.60A were evaluated in BRET2, PLA, and bind-ing assays. The BRET2 signal of the A2AAR-D2R interactionwas significantlyreduced for both mutants. In addition, A2AAR agonist CGS-21680 reduced theaffinity of D2R agonist Quinpirole for both the L2075.56A / K2115.60A D2Rand Y1925.41A D2R mutants, but not to the same extent as for the D2R wild-type. This reduction in response to the negative allosteric modulation supportsa role of these residues in the dimer interface and allosteric modulation.

Attempts to further characterize the dimer interface by studying its speci-ficity are still ongoing. Comparisons between A2AAR and A1AR, which doesnot dimerize with D2R, suggested potential mutation sites to understand how

44

Figure 8.2. Model the A2AAR-D2R heterodimer. The residues selected for the muta-genesis study are indicated with black rectangles.

the A2AAR-D2R heterodimer is formed. Moreover, a heteromer comprisingA2AAR, D2R and mGluR5 has also been observed and we have also investi-gated the structure of this complex [149, 150].

In Paper IV we integrated biophysical experiments with computational ap-proaches to characterize the A2AAR-D2R heterodimer. BRET experimentsallowed us to derive the TM helices involved in the interface. By combiningMD simulations, BRET data and mutagenesis we were able to derive potentialhotspots in the A2AAR-D2R heterodimer interface.

45

9. Conclusions and perspectives

The past decades have seen several breakthroughs in GPCR signaling, phar-macology and structural biology. These discoveries resulted in a series of newparadigms in GPCR structure-based drug discovery [45]. The work in thisthesis has focused on how structural information can be used to predict thestructures of GPCRs, GPCR-drug complexes and GPCR heterodimer inter-faces. We investigated the potential of SBVS methods for the rational designof a novel generation of GPCR therapeutics.

With the current wealth of atomic-level information, HM can be used topredict the majority of GPCR structures [151]. However, a few key resultsemerged from Paper I. A suitable description of template quality cannot belimited to a single model performance but should rather be evaluated based onmedian enrichment of an ensemble of models. Moreover, the use of a set ofmodels to derive an ensemble enrichment can be a powerful approach to in-corporate receptor flexibility, which is expected to promote the identificationof a diverse set of chemotypes [152] and thus improve hit rates. For challeng-ing cases such as deorphanization of receptors, the lack of close templates stillhinders the identification of putative ligands. Based on the median enrichmentand ligand binding modes obtained for remote templates, usage of distant tem-plates is not recommended in GPCR modeling. However, the use of differenttemplates with good local similarity for distinct parts of the receptor couldcontribute to better models. Finally, results in Paper I showed that MD sim-ulations do not necessarily improve homology model performances in SBVS.Nonetheless, MD simulations in combination with VS can be an alternativemeans to incorporate flexibility in the docking calculations [152, 153].

In Papers II and III, we demonstrated that when experimental structuresor validated models are available, SBVS can be successfully used to identifyligands with tailored properties, e.g. multi-target and functionally selectiveligands. However, more structures are still needed. A rigorous study of func-tional selectivity will require experimental structures of different active statesbound to various biased ligands and use of MD simulations to derive insightfulSAR. Moreover, rational optimization of multi-target ligands will necessitatenew strategies such as incorporation of BS flexibility or the use of tailoredmade libraries.

46

The combined use of biophysical and computational approaches in PaperIV allowed the elucidation of a GPCR heterodimer interface. The increas-ing number of GPCR oligomer structures thanks to advances in crystallizationand cryo-EM is contributing to the structural and functional understanding ofGPCR quaternary structure. Moreover, this structural information, e.g com-mon interactions observed in protein-protein complexes, could guide and con-tribute to the computational prediction of GPCR interactome.

47

10. Sammanfattning

G-proteinkopplade receptorer (GPCRer) utgör en stor grupp av membran-proteiner som spelar nyckelroller i många fysiologiska processer. När krop-pens signalsubstanser aktiverar en receptor överförs signaler från utsidan avcellen till insidan. Förändringar i receptorers signalering kan leda till sjuk-domar i centrala nervsystemet, hjärt-kärlsjukdomar och cancer. På grundav detta är GPCRer viktiga mål för läkemedelsutveckling. För närvarandeverkar 34% av alla läkemedel genom att interagera med GPCRer. Under desenaste åren har forskning resulterat i en djupare förståelse för GPCRers funk-tion och flera vetenskapliga genombrott har gjort det möjligt att bestämmaderas tredimensionella strukturer. Strukturerna kan användas för att försökaförstå hur GPCRer fungerar på molekylär nivå samt för att identifiera nyaläkemedelskandidater med datorberäkningar. Nuvarande metoder för att iden-tifiera läkemedelskandidater för behandling av komplexa sjukdomar är otill-räckliga och kostnadskrävande. Dagens kraftfulla datorer ger möjligheter attanvända beräkningar för att göra metoderna effektivare. Denna avhandlingfokuserar på hur den stora mängden strukturell data som samlats in under desenaste åren kan användas för att modellera GPCRer på molekylär nivå och föratt identifiera läkemedelskandidater med skräddarsydda egenskaper. I artikel(I) utvecklades nya metoder för att förutsäga GPCRers struktur med hjälp avdatormodeller. Vi utvärderade om dessa modeller var lämpliga för använd-ning i läkemedelsutveckling. I artikel (II) användes datormodeller för att sökaigenom stora databaser med substanser för att hitta sådana som kan interageramed både A2A-adenosinreceptorn och monoaminoxidas B. Flera intressantasubstanser identifierades, vilket kan bidra till nya läkemedel mot Parkinsonssjukdom. På liknande sätt identifierades i artikel (III) molekyler som aktiverarD2-dopaminreceptorn på ett sätt som kan leda till läkemedel med färre bief-fekter. Resultaten i artiklarna (II) och (III) visar att strukturbaserade metoderkan bidra till att identifiera läkemedelskandidater med skräddarsydda egen-skaper. I artikel (IV) undersöktes hur strukturell information, beräkningsme-toder och biofysikaliska metoder kan användas för att modellera interaktionermellan GPCRer i cellens membran. Sådana interaktioner har visat sig påverkaGPCRers aktivitet och representerar ett helt nytt koncept för hur signaleringregleras. Vår modell av komplexet mellan adenosin- och dopamin-receptorerkunde bidra till att förstå dessa interaktioner på molekylär nivå.

48

11. Acknowledgments

To Isatou, my mom: I dedicate this thesis to you. I am deeply grateful foryour unconditional support, trust and faith in me. Although it took some timeto grasp what ”djon bè ni da kan”means to me, you were always there. You’vetaught me the importance and value of education and investment in oneself.You’ve also taught me one of the most important thing in life: how to read.You are my most beloved and respected teacher. Allahma i simaya la.

To my brothers and sisters Abdoula, Fatou, Zac, and Ami: merci à tous pourvotre soutien, écoute et fraternité. Parler de tout et de rien, m’a permis degarder les pieds sur terre. Merci Abdoula, pour la couverture. Tu déchires !

Most of what I’ve learned during my Ph.D. did not come from literature butfrom fruitful discussions with my colleagues and peers. I would like to expressmy gratitude:

To Jens: thank you for taking a chance on me and giving me the opportunityto start this long dive into membrane proteins. Your constant enthusiasm, sup-port, and patience was invaluable. Discussing with you on all the stimulatingprojects has taught me a lot and fostered my curiosity. I have learned whatresearch means through the numerous meetings and discussions we had. Forthat, I am truly grateful.

To David van der Spoel: thank you very much for your steady support assecond supervisor. Changing department is not easy and you helped us ac-commodate smoothly in this new environment.

The current and former members of the Carlsson Lab have made this Ph.D.one of the best experience in my life.

To Pierre: thank you for your friendship. We’ve joined the group roughly atthe same time and yet we’ve apparently formed The French mafia. Thank youfor all the fruitful discussions, for helping me with food, for sharing your ex-periences.

To Flavio: you still owe me a table tennis game. Thank you so much for allthe talks about science, society, guitar, food, and Italy.

49

To Duy: thank you for taking me as a lab intern. You will certainly be a greatchemistry teacher. Thank you so much for showing me how compounds aremade and giving me the chance to learn a bit of synthetic chemistry. Can I sayI can make compounds?

ToNicolas: another member of the French mafia. Thank you so much for yoursupport, all the advice and discussions about living green.

To Stefanie: thank you for your warm support. I think our office has the bestambient temperature. I wish you a good luck with your Ph.D.

ToAndreas: thanks for the enriching discussions about docking, teaching andall. Good luck with your Ph.D.

To Jon: thanks for all the discussions about coding, MD and other geeky stuffs.

ToWilber and Marcus: thank you so much for letting me play with cell cul-tures and gels. Honestly, I think I just like the colors in cell imaging ! AlthoughI cannot claim any expertise, I have learned valuable things about biologicaland biophysical experiments. Thank you so much.

To David, Alexey, Anirudh and Axel: thank you very much. Most of whatI’ve learned come from the many discussions we had. I have learned from youwhat one needs to dive into the world of GPCRs. Thank you for sharing yourknowledge and wisdom about research and Ph.D. studies. Thank you for yourstory telling and gossiping too.

To the members of the GPCR Book Club: thanks for the fruitful discussionsand the knowledge sharing.

To Tex: thanks for sharing teas.

To the ICM, teaching and Ph.D. studies admins: thank you for your patienceand continuous support.

To the people at ScilifeLab, aplha6: thank you so much for welcoming usand introducing us to the concept of ”Fika” seen through the eyes of interna-tional students and researchers.

To Ozge: thank you for being my Drama mate. We definitely need to visitKorea and Japan together.

50

To people at Orgfarm: thank you so much for welcoming us and giving methe opportunity to do teaching. It was an invaluable experience.

Moving to a new country is often scary and require some adaptation. Life isfull of important meetings and I have met extraordinary people. Truly, it is thepeople who make one love a place.

To Samantha and Lo: thank you for being my adoptive family in Stockholm.

To Najla and Aminata: merci infiniment pour tous les délires, les voyages,votre soutien et vos encouragements.

To Ece: thank you for being my big sister. Living with you has given me asense of peace and belonging. Thank you for all the support, the talks and ev-erything.

To Sammy: you are the first friend I made in Sweden. I have discovered thebeauty of Stockholm with you. Thank you for all the talks and experiences weshared.

To Sonia, Emmeline, Ali, and Riad, my dearest friends: thank you for yourunwavering support and friendship throughout all these years.

51

References

[1] Langley, J. N. On the reaction of cells and of nerve-endings to certain poisons,chiefly as regards the reaction of striated muscle to nicotine and to curari. J.Physiol. 33, 374–413 (1905).

[2] Ehrlich, P. Partial cell functions: Nobel lecture, December 11, 1908. InPhysiol. or Med. Incl. Present. speeches laureates’ Biogr.; 1901-1921,304–321 (1908).

[3] Dixon, R. A. et al. Cloning of the gene and cDNA for mammalianβ-adrenergic receptor and homology with rhodopsin. Nature 321, 75–79(1986).

[4] Lefkowitz, R. J. A brief history of G-protein coupled receptors (NobelLecture). Angew. Chemie - Int. Ed. 52, 6366–6378 (2013).

[5] Fredriksson, R., Lagerström, M. C., Lundin, L. G. & Schiöth, H. B. TheG-protein-coupled receptors in the human genome form five main families.Phylogenetic analysis, paralogon groups, and fingerprints. Mol. Pharmacol.63, 1256–1272 (2003).

[6] Hermans, E. Biochemical and pharmacological control of the multiplicity ofcoupling at G-protein-coupled receptors. Pharmacol. Ther. 99, 25–44 (2003).

[7] Bridges, T. M. & Lindsley, C. W. G-protein-coupled receptors: From classicalmodes of modulation to allosteric mechanisms. ACS Chem. Biol. 3, 530–541(2008).

[8] Marinissen, M. J. & Gutkind, J. S. G-protein-coupled receptors and signalingnetworks: Emerging paradigms. Trends Pharmacol. Sci. 22, 368–376 (2001).

[9] Ritter, S. L. & Hall, R. A. Fine-tuning of GPCR activity byreceptor-interacting proteins. Nat. Rev. Mol. Cell Biol. 10, 819–30 (2009).

[10] Smith, J. S., Lefkowitz, R. J. & Rajagopal, S. Biased signalling: From simpleswitches to allosteric microprocessors. Nat. Rev. Drug Discov. 17, 243–260(2018).

[11] Oldham, W. M. & Hamm, H. E. Structural basis of function in heterotrimericG proteins. Q. Rev. Biophys. 39, 117–166 (2006).

52

[12] Krishna Kumar, K. et al. Structure of a Signaling Cannabinoid Receptor 1-GProtein Complex. Cell 176, 448–458 (2019).

[13] Yin, W. et al. A complex structure of arrestin-2 bound to a G protein-coupledreceptor. Cell Res. 29, 971–983 (2019).

[14] Wacker, D., Stevens, R. C. & Roth, B. L. How Ligands Illuminate GPCRMolecular Pharmacology. Cell 170, 414–427 (2017).

[15] Reiter, E., Ahn, S., Shukla, A. K. & Lefkowitz, R. J. Molecular Mechanism ofβ-Arrestin-Biased Agonism at Seven-Transmembrane Receptors. Annu. Rev.Pharmacol. Toxicol. 52, 179–197 (2012).

[16] Liu, J. J., Horst, R., Katritch, V., Stevens, R. C. & Wüthrich, K. Biasedsignaling pathways in β2-adrenergic receptor characterized by 19F-NMR.Science 335, 1106–1110 (2012).

[17] Kahsai, A. W. et al. Multiple ligand-specific conformations of theβ2-adrenergic receptor. Nat. Chem. Biol. 7, 692–700 (2011).

[18] Chen, X. T. et al. Structure-activity relationships and discovery of a G proteinbiased µ opioid receptor ligand, [(3-methoxythiophen-2-yl)methyl]({2-[(9r)-9-(pyridin-2-yl)-6-oxaspiro-[4.5]decan-9-yl]ethyl})amine (TRV130), for thetreatment of acute severe pain. J. Med. Chem. 56, 8019–8031 (2013).

[19] Violin, J. D., Crombie, A. L., Soergel, D. G. & Lark, M. W. Biased ligands atG-protein-coupled receptors: Promise and progress. Trends Pharmacol. Sci.35, 308–316 (2014).

[20] May, L. T., Leach, K., Sexton, P. M. & Christopoulos, A. AllostericModulation of G Protein–Coupled Receptors. Annu. Rev. Pharmacol. Toxicol.47, 1–51 (2007).

[21] Christopoulos, A. & Kenakin, T. G protein-coupled receptor allosterism andcomplexing. Pharmacol. Rev. 54, 323–374 (2002).

[22] Christopoulos, A. Allosteric binding sites on cell-surface receptors: Noveltargets for drug discovery. Nat. Rev. Drug Discov. 1, 198–210 (2002).

[23] Smith, N. J. & Milligan, G. Allostery at G Protein-Coupled Receptor Homo-and Heteromers: Uncharted Pharmacological Landscapes. Pharmacol. Rev.62, 701–725 (2010).

[24] Marshall, F. H., Jones, K. A., Kaupmann, K. & Bettler, B. GABAB receptors -The first 7TM heterodimers. Trends Pharmacol. Sci. 20, 396–399 (1999).

53

[25] Möhler, H. & Fritschy, J. M. GABAB receptors make it to the top - As dimers.In Trends Pharmacol. Sci., vol. 20, 87–89 (1999).

[26] Fung, J. J. et al. Ligand-regulated oligomerization of β2-adrenoceptors in amodel lipid bilayer. EMBO J. 28, 3315–3328 (2009).

[27] Franco, R., Martínez-Pinilla, E., Lanciego, J. L. & Navarro, G. BasicPharmacological and Structural Evidence for Class A G-Protein-CoupledReceptor Heteromerization. Front. Pharmacol. 7, 76 (2016).

[28] Milligan, G. G Protein-Coupled Receptor Dimerization: Function and LigandPharmacology. Mol. Pharmacol. 66, 1–7 (2004).

[29] Ferré, S. et al. G protein-coupled receptor oligomerization revisited:Functional and pharmacological perspectives. Pharmacol. Rev. 66, 413–434(2014).

[30] Felce, J. H. et al. Receptor Quaternary Organization Explains GProtein-Coupled Receptor Family Structure. Cell Rep. 20, 2654–2665 (2017).

[31] Isberg, V. et al. GPCRdb: An information system for G protein-coupledreceptors. Nucleic Acids Res. 44, D356–D364 (2016).

[32] Shimada, I., Ueda, T., Kofuku, Y., Eddy, M. T. & Wüthrich, K. GPCR drugdiscovery: integrating solution NMR data with crystal and cryo-EM structures.Nat. Rev. Drug Discov. 18, 59–82 (2018).

[33] Palczewski, K. et al. Crystal Structure of Rhodopsin: A G Protein-CoupledReceptor. Science 289, 739–745 (2000).

[34] Rasmussen, S. G. et al. Crystal structure of the human β2 adrenergicG-protein-coupled receptor. Nature 450, 383–387 (2007).

[35] Ghosh, E., Kumari, P., Jaiman, D. & Shukla, A. K. Methodological advances:the unsung heroes of the GPCR structural revolution. Nat. Rev. Mol. Cell Biol.16, 69–81 (2015).

[36] Cherezov, V. et al. High-resolution crystal structure of an engineered humanβ2-adrenergic G protein-coupled receptor. Science 318, 1258–1265 (2007).

[37] Warne, T. et al. Structure of a β1-adrenergic G-protein-coupled receptor.Nature 454, 486–491 (2008).

[38] Doré, A. S. et al. Structure of the adenosine A2A receptor in complex withZM241385 and the xanthines XAC and caffeine. Structure 19, 1283–1293(2011).

54

[39] Ring, A. M. et al. Adrenaline-activated structure of β2-adrenoceptor stabilizedby an engineered nanobody. Nature 502, 575–579 (2013).

[40] Serrano-Vega, M. J., Magnani, F., Shibata, Y. & Tate, C. G. Conformationalthermostabilization of the β1-adrenergic receptor in a detergent-resistant form.Proc. Natl. Acad. Sci. U. S. A. 105, 877–882 (2008).

[41] Liang, Y. L. et al. Phase-plate cryo-EM structure of a class B GPCR-G-proteincomplex. Nature 546, 118–123 (2017).

[42] García-Nafría, J. & Tate, C. G. Cryo-EM structures of GPCRs coupled to Gs,Gi and Go. Mol. Cell. Endocrinol. 488, 1–13 (2019).

[43] Heng, B. C., Aubel, D. & Fussenegger, M. An overview of the diverse roles ofG-protein coupled receptors (GPCRs) in the pathophysiology of varioushuman diseases. Biotechnol. Adv. 31, 1676–1694 (2013).

[44] Hauser, A. S., Attwood, M. M., Rask-Andersen, M., Schiöth, H. B. & Gloriam,D. E. Trends in GPCR drug discovery: new agents, targets and indications.Nat. Rev. Drug Discov. 16, 829–842 (2017).

[45] Jacobson, K. A. New paradigms in GPCR drug discovery. Biochem.Pharmacol. 98, 541–555 (2015).

[46] Roth, B. L., Sheffler, D. J. & Kroeze, W. K. Magic shotguns versus magicbullets: selectively non-selective drugs for mood disorders and schizophrenia.Nat. Rev. Drug Discov. 3, 353–359 (2004).

[47] Anighoro, A., Bajorath, J. & Rastelli, G. Polypharmacology: Challenges andopportunities in drug discovery. J. Med. Chem. 57, 7874–7887 (2014).

[48] Rankovic, Z., Brust, T. F. & Bohn, L. M. Biased agonism: An emergingparadigm in GPCR drug discovery. Bioorganic Med. Chem. Lett. 26, 241–250(2016).

[49] Craig Venter, J. et al. The sequence of the human genome. Science 291,1304–1351 (2001).

[50] Lander, S. et al. Initial sequencing and analysis of the human genome. Nature409, 860–921 (2001).

[51] Pertea, M. & Salzberg, S. L. Between a chicken and a grape: estimating thenumber of human genes. Genome Biol. 11, 206 (2010).

[52] The UniProt Consortium. UniProt: A worldwide hub of protein knowledge.Nucleic Acids Research 47, D506–D515 (2019).

55

[53] Baker, D. & Sali, A. Protein structure prediction and structural genomics.Science 294, 93–96 (2001).

[54] Michino, M. et al. Community-wide assessment of GPCR structure modellingand ligand docking: GPCR Dock 2008. Nat. Rev. Drug Discov. 8, 455–463(2009).

[55] Kufareva, I. et al. Status of GPCR modeling and docking as reflected bycommunity-wide GPCR Dock 2010 assessment. Structure 19, 1108–1126(2011).

[56] Kufareva, I. et al. Advances in GPCR modeling evaluated by the GPCR dock2013 assessment: Meeting new challenges. Structure 22, 1120–1139 (2014).

[57] Nakashima, H., Nishikawa, K. & Ooi, T. The folding type of a protein isrelevant to the amino acid composition. J. Biochem. 99, 153–162 (1986).

[58] Jones, D. T. Progress in protein structure prediction. Curr. Opin. Struct. Biol.7, 377–387 (1997).

[59] Bonneau, R. & Baker, D. Ab initio protein structure prediction: Progress andprospects. Annu. Rev. Biophys. Biomol. Struct. 30, 173–189 (2001).

[60] Vaidehi, N. et al. Prediction of structure and function of G protein-coupledreceptors. Proc. Natl. Acad. Sci. U. S. A. 99, 12622–12627 (2002).

[61] Shacham, S. et al. Modeling the 3D structure of GPCRs from sequence. Med.Res. Rev. 21, 472–483 (2001).

[62] Spijker, P., Vaidehi, N., Freddolino, P. L., Hilbers, P. A. & Goddard, W. A.Dynamic behavior of fully solvated β2-adrenergic receptor, embedded in themembrane with bound agonist or antagonist. Proc. Natl. Acad. Sci. U. S. A.103, 4882–4887 (2006).

[63] Greener, J. G., Kandathil, S. M. & Jones, D. T. Deep learning extends de novoprotein modelling coverage of genomes using iteratively predicted structuralconstraints. Nat. Commun. 10, 3977 (2019).

[64] Jones, D. T., Buchan, D. W., Cozzetto, D. & Pontil, M. PSICOV: Precisestructural contact prediction using sparse inverse covariance estimation onlarge multiple sequence alignments. Bioinformatics 28, 184–190 (2012).

[65] Marks, D. S. et al. Protein 3D structure computed from evolutionary sequencevariation. PLoS One 6, e28766 (2011).

[66] Evans, R. et al. De novo structure prediction with deep-learning based scoring.Tech. Rep. (2018).

56

[67] AlQuraishi, M. AlphaFold at CASP13. Bioinformatics 1–4 (2019).

[68] Schaeffer, R. D. & Daggett, V. Protein folds and protein folding. Protein Eng.Des. Sel. 24, 11–19 (2011).

[69] Sadreyev, R. I., Kim, B. H. & Grishin, N. V. Discrete-continuous duality ofprotein structure space. Curr. Opin. Struct. Biol. 19, 321–328 (2009).

[70] Zhang, Y., Devries, M. E. & Skolnick, J. Structure modeling of all identified Gprotein-coupled receptors in the human genome. PLoS Comput. Biol. 2, 88–99(2006).

[71] Chothia, C. & Lesk, A. M. The relation between the divergence of sequenceand structure in proteins. EMBO J. 5, 823–6 (1986).

[72] Krieger, E., Nabuurs, S. B. & Vriend, G. Homology Modeling. InMethodsMol. Biol., vol. 44, 507–521 (2003).

[73] Rost, B. Twilight zone of protein sequence alignments. Protein Eng. 12,85–94 (1999).

[74] Jaiteh, M., Taly, A. & Hénin, J. Evolution of Pentameric Ligand-Gated IonChannels: Pro-Loop Receptors. Plos One 11, e0151934 (2016).

[75] Ballesteros, J. A. & Weinstein, H. Integrated methods for the construction ofthree-dimensional models and computational probing of structure-functionrelations in G protein-coupled receptors. Methods Neurosci. 25, 366–428(1995).

[76] Isberg, V. et al. Generic GPCR residue numbers - Aligning topology mapswhile minding the gaps. Trends Pharmacol. Sci. 36, 22–31 (2015).

[77] Venkatakrishnan, A. J. et al. Molecular signatures of G-protein-coupledreceptors. Nature 494, 185–194 (2013).

[78] Laskowski, R. A., MacArthur, M. W., Moss, D. S. & Thornton, J. M.PROCHECK: a program to check the stereochemical quality of proteinstructures. J. Appl. Crystallogr. 26, 283–291 (1993).

[79] Shen, M.-Y. & Sali, A. Statistical potential for assessment and prediction ofprotein structures. Protein Science 15, 2507–2524 (2006).

[80] Carlsson, J. et al. Ligand discovery from a dopamine D3 receptor homologymodel and crystal structure. Nat. Chem. Biol. 7, 769–778 (2011).

[81] Katritch, V., Kufareva, I. & Abagyan, R. Structure based prediction of

57

subtype-selectivity for adenosine receptor antagonists. Neuropharmacology60, 108–115 (2011).

[82] Šali, A. & Blundell, T. L. Comparative protein modelling by satisfaction ofspatial restraints. J. Mol. Biol. 234, 779–815 (1993).

[83] Kühhorn, J., Hübner, H. & Gmeiner, P. Bivalent dopamine D2 receptor ligands:Synthesis and binding properties. J. Med. Chem. 54, 4896–4903 (2011).

[84] Schonenbach, N. S., Hussain, S. & O’Malley, M. A. Structure and function ofG protein-coupled receptor oligomers: Implications for drug discovery. WileyInterdiscip. Rev. Nanomedicine Nanobiotechnology 7, 408–427 (2015).

[85] Huang, S. Y. Search strategies and evaluation in protein-protein docking:Principles, advances and challenges. Drug Discov. Today 19, 1081–1096(2014).

[86] Méndez, R., Leplae, R., Lensink, M. F. & Wodak, S. J. Assessment of CAPRIpredictions in rounds 3-5 shows progress in docking procedures. ProteinsStruct. Funct. Bioinforma. 60, 150–169 (2005).

[87] Dominguez, C., Boelens, R. & Bonvin, A. M. HADDOCK: A protein-proteindocking approach based on biochemical or biophysical information. J. Am.Chem. Soc. 125, 1731–1737 (2003).

[88] Janin, J. et al. CAPRI: A critical assessment of PRedicted interactions.Proteins Struct. Funct. Genet. 52, 2–9 (2003).

[89] Lensink, M. F., Méndez, R. & Wodak, S. J. Docking and scoring proteincomplexes: CAPRI 3rd Edition. Proteins Struct. Funct. Bioinforma. 69,704–718 (2007).

[90] Lensink, M. F. & Wodak, S. J. Blind predictions of protein interfaces bydocking calculations in CAPRI. Proteins Struct. Funct. Bioinforma. 78,3085–3095 (2010).

[91] Lensink, M. F. & Wodak, S. J. Docking, scoring, and affinity prediction inCAPRI. Proteins Struct. Funct. Bioinforma. 81, 2082–2095 (2013).

[92] Lensink, M. F., Velankar, S. & Wodak, S. J. Modeling protein–protein andprotein–peptide complexes: CAPRI 6th edition. Proteins Struct. Funct.Bioinforma. 85, 359–377 (2017).

[93] Koshland, D. E. Application of a Theory of Enzyme Specificity to ProteinSynthesis. Proc. Natl. Acad. Sci. 44, 98–104 (1958).

58

[94] Ma, B., Shatsky, M., Wolfson, H. J. & Nussinov, R. Multiple diverse ligandsbinding at a single protein site: A matter of pre-existing populations. ProteinSci. 11, 184–197 (2009).

[95] Latorraca, N. R., Venkatakrishnan, A. J. & Dror, R. O. GPCR dynamics:Structures in motion. Chem. Rev. 117, 139–155 (2017).

[96] Bostock, M. J., Solt, A. S. & Nietlispach, D. The role of NMR spectroscopy inmapping the conformational landscape of GPCRs. Curr. Opin. Struct. Biol. 57,145–156 (2019).

[97] Nango, E. et al. A three-dimensionalmovie of structural changes inbacteriorhodopsin. Science 354, 1552–1557 (2016).

[98] Karplus, M., McCammon, J. A. & Gelin, B. R. Dynamics of folded proteins.Nature 267, 585–590 (1977).

[99] Karplus, M. & McCammon, J. A. Molecular dynamics simulations ofbiomolecules. Nat. Struct. Biol. 9, 646–652 (2002).

[100] MacKerell, A. D. et al. All-atom empirical potential for molecular modelingand dynamics studies of proteins. J. Phys. Chem. B 102, 3586–3616 (1998).

[101] Jorgensen, W. L., Maxwell, D. S. & Tirado-Rives, J. Development and testingof the OPLS all-atom force field on conformational energetics and propertiesof organic liquids. J. Am. Chem. Soc. 118, 11225–11236 (1996).

[102] Bayly, C. I. et al. A Second Generation Force Field for the Simulation ofProteins, Nucleic Acids, and Organic Molecules. J. Am. Chem. Soc. 117,5179–5197 (1995).

[103] Harvey, M. J., Giupponi, G. & De Fabritiis, G. ACEMD: Acceleratingbiomolecular dynamics in the microsecond time scale. J. Chem. TheoryComput. 5, 1632–1639 (2009).

[104] Essmann, U. et al. A smooth particle mesh Ewald method. J. Chem. Phys.103, 8577–8593 (1995).

[105] Pronk, S. et al. GROMACS 4.5: A high-throughput and highly parallel opensource molecular simulation toolkit. Bioinformatics 29, 845–854 (2013).

[106] Verlet, L. Computer ”experiments” on classical fluids. I. Thermodynamicalproperties of Lennard-Jones molecules. Phys. Rev. 159, 98–103 (1967).

[107] Kuntz, I. D., Meng, E. C. & Shoichet, B. K. Structure-Based MolecularDesign. Acc. Chem. Res. 27, 117–123 (1994).

59

[108] Chen, L. et al. From Laptop to Benchtop to Bedside: Structure-based DrugDesign on Protein Targets. Curr. Pharm. Des. 18, 1217–1239 (2012).

[109] Irwin, J. J. & Shoichet, B. K. Docking Screens for Novel Ligands ConferringNew Biology. J. Med. Chem. 59, 4103–4120 (2016).

[110] Huang, N. & Shoichet, B. K. Exploiting ordered waters in molecular docking.J. Med. Chem. 51, 4862–4865 (2008).

[111] Irwin, J. J. & Shoichet, B. K. ZINC − A Free Database of CommerciallyAvailable Compounds for Virtual Screening. J. Chem. Inf. Model 45, 177–182(2005).

[112] Baell, J. B. & Holloway, G. A. New substructure filters for removal of panassay interference compounds (PAINS) from screening libraries and for theirexclusion in bioassays. J. Med. Chem. 53, 2719–2740 (2010).

[113] Welsch, M. E., Snyder, S. A. & Stockwell, B. R. Privileged scaffolds forlibrary design and drug discovery. Curr. Opin. Chem. Biol. 14, 347–361(2010).

[114] Kitchen, D., Decornez, H., Furr, J. & Bajorath, J. Docking and scoring invirtual screening for drug discovery: methods and applications. Nat. Rev. DrugDiscov. 3, 935–949 (2004).

[115] Böhm, H. J. LUDI: rule-based automatic design of new substituents forenzyme inhibitor leads. J. Comput. Aided. Mol. Des. 6, 593–606 (1992).

[116] Muegge, I. & Martin, Y. C. A general and fast scoring function forprotein-ligand interactions: A simplified potential approach. J. Med. Chem.42, 791–804 (1999).

[117] Lorber, D. M. & Shoichet, B. K. Hierarchical docking of databases of multipleligand conformations. Curr. Top. Med. Chem. 5, 739–49 (2005).

[118] Bento, A. P. et al. The ChEMBL bioactivity database: An update. NucleicAcids Research 42, D1083–D1090 (2014).

[119] Mysinger, M. M., Carchia, M., Irwin, J. J. & Shoichet, B. K. Directory ofuseful decoys, enhanced (DUD-E): Better ligands and decoys for betterbenchmarking. J. Med. Chem. 55, 6582–6594 (2012).

[120] Mysinger, M. M. & Shoichet, B. K. Rapid context-dependent liganddesolvation in molecular docking. J. Chem. Inf. Model. 50, 1561–1573 (2010).

[121] Liang, Y. L. et al. Phase-plate cryo-EM structure of a biased agonistboundhuman GLP-1 receptor-Gs complex. Nature 555, 121–125 (2018).

60

2

2A

Chen, J. F. et al. 8-(3-Chlorostyryl) caffeine may attenuate MPTP neurotoxicity through dual actions of monoamine oxidase inhibition and A2A receptor antagonism. J Biol. Chem. 277, 36040–36044 (2002).

[136] Armentero, M. T. et al. Past, present and future of A2A adenosine receptorantagonists in the therapy of Parkinson’s disease. Pharmacol. Ther. 132,280–299 (2011).

[137] Riederer, P. & Laux, G. MAO-inhibitors in Parkinson’s Disease. Exp.Neurobiol. 20, 1–17 (2011).

[138] Binda, C. et al. Structures of human monoamine oxidase B complexes withselective noncovalent inhibitors: Safinamide and coumarin analogs. J. Med.Chem. 50, 5848–5852 (2007).

[139] Barelier, S., Sterling, T., O’Meara, M. J. & Shoichet, B. K. The Recognition ofIdentical Ligands by Unrelated Proteins. ACS Chem. Biol. 10, 2772–2784(2015).

[140] Besnard, J. et al. Automated design of ligands to polypharmacologicalprofiles. Nature 492, 215–20 (2012).

[141] Wang, Y. et al. PubChem BioAssay: 2017 update. Nucleic Acids Res. 45,D955–D963 (2017).

[142] Sun, B. et al. Crystal structure of the adenosine A2A receptor bound to anantagonist reveals a potential allosteric pocket. Proc. Natl. Acad. Sci. 114,2066–2071 (2017).

[143] Bock, A., Kostenis, E., Tränkle, C., Lohse, M. J. & Mohr, K. Pilot the pulse:Controlling the multiplicity of receptor dynamics. Trends Pharmacol. Sci. 35,630–638 (2014).

[144] Hilger, D., Masureel, M. & Kobilka, B. K. Structure and dynamics of GPCRsignaling complexes. Nat. Struct. Mol. Biol. 25, 4–12 (2018).

[145] Svenningsson, P., Le Moine, C., Fisone, G. & Fredholm, B. B. Distribution,biochemistry and function of striatal adenosine A2A receptors. Prog.Neurobiol. 59, 355–396 (1999).

[146] Fernández-Ruiz, J., Hernández, M. & Ramos, J. A. Cannabinoid-dopamineinteraction in the pathophysiology and treatment of CNS disorders. CNSNeurosci. Ther. 16, e18–e42 (2010).

[147] Ferre, S., Von Euler, G., Johansson, B., Fredholm, B. B. & Fuxe, K.Stimulation of high-affinity adenosine A2 receptors decreases the affinity ofdopamine D2 receptors in rat striatal membranes. Proc. Natl. Acad. Sci. U. S.A. 88, 7238–7241 (1991).

[148] Borroto-Escuela, D. O. et al. Characterization of the A2AR-D2R interface:Focus on the role of the C-terminal tail and the transmembrane helices.

62

Biochem. Biophys. Res. Commun. 402, 801–807 (2010).

[149] Cabello, N. et al. Metabotropic glutamate type 5, dopamine D2 and adenosineA2A receptors form higher-order oligomers in living cells. J. Neurochem. 109,1497–1507 (2009).

[150] Fuxe, K. et al. Receptor heteromerization in adenosine A2A receptorsignaling: Relevance for striatal function and Parkinson’s disease. Neurology61, S19–S23 (2003).

[151] Pándy-Szekeres, G. et al. GPCRdb in 2018: Adding GPCR structure modelsand ligands. Nucleic Acids Res. 46, D440–D446 (2018).

[152] Ellingson, S. R., Miao, Y., Baudry, J. & Smith, J. C. Multi-ConformerEnsemble Docking to Difficult Protein Targets. J. Phys. Chem. B 119,1026–1034 (2015).

[153] Tarcsay, A. et al. The impact of molecular dynamics sampling on theperformance of virtual screening against GPCRs. J. Chem. Inf. Model. 53,2990–9 (2013).

63

Acta Universitatis UpsaliensisDigital Comprehensive Summaries of Uppsala Dissertationsfrom the Faculty of Science and Technology 1889

Editor: The Dean of the Faculty of Science and Technology

A doctoral dissertation from the Faculty of Science andTechnology, Uppsala University, is usually a summary of anumber of papers. A few copies of the complete dissertationare kept at major Swedish research libraries, while thesummary alone is distributed internationally throughthe series Digital Comprehensive Summaries of UppsalaDissertations from the Faculty of Science and Technology.(Prior to January, 2005, the series was published under thetitle “Comprehensive Summaries of Uppsala Dissertationsfrom the Faculty of Science and Technology”.)

Distribution: publications.uu.seurn:nbn:se:uu:diva-399133

ACTAUNIVERSITATIS

UPSALIENSISUPPSALA

2020