Homology modeling of adenosine A3 receptor

Kamil J. Kuder , Peter Kolb , Katarzyna Kie-Kononowicz

a

Department of Technology and Biotechnology of Drugs,

Jagiellonian University Medical College,
Faculty of Pharmacy, Medyczna 9, 30-688 Krakw, Poland;
b
Department of Pharmaceutical Chemistry, Philipps-Universitt Marburg,
Wilhelm-Roser-Strae 2, Marburg, Germany
a

VIII Konwersatorium
Chemii Medycznej
Lublin
15-17.09 2016

kamil.kuder@uj.edu.pl
Introduction

Model optimization and validation

Adenosine receptor ligands are currently being developed as promising agents for CNS
disorders (Parkinsons, Alzheimers diseases, ischaemia) [1]. Also, the adenosine A2A receptor
has been co-crystallized with several ligands: agonists as well as antagonists. It thus serves
as an optimal template with a well-described orthosteric ligand binding region for AR.
In order to investigate the usability of homology models in this context, multiple adenosine
A1 receptor (A1R) homology models have been previously obtained and a library of lead-like
compounds has been docked [2]. As a result, a number of potent and a few selective ligands
toward the intended target were found. However, in in vitro experimental verification studies
many ligands also bound to A2AAR and A3AR. Therefore, the aim of this work was to build an
A3AR homology model and validate the usability ofthus obtained receptor for further docking
studies.

The next step was to test the enrichment of the binding pockets of three selected models of
binders over non-binders. For this purpose, two sets of ligands were obtained from
ChEMBL database [11]. Binders set consisted of ca. 1500 molecules described as A3AR
ligands with Ki 100nM. Second, non-binders set consisted of ca. 800 molecules, tested
for A3AR and described as inactive for this target. Structures of both sets of ligands were
obtained from ZINC database, by searching for corresponding ZINC ID's for all of the
ligands extracted from ChEMBL. High quality 3D conformer ensembles of both sets were
obtained using OMEGA module of OEDocking software package (maximum No. of
conformers = 100; RMS = 0.5) [12].
As a reference ligand for docking, co-crystallized within 3EML structure, the ZMA ligand
was placed in the A3 target receptor models after their alignment to 3EML structure. Two
sets of ligands were then docked to the prepared receptor homology models using
HYBRID module (one pose per ligand, max. hitlist size - 500 molecules), implemented in
OEDocking Software. After docking, the top 500 poses were inspected visually, and
receiver operator characteristic (ROC) curves were generated along with calculation of the
area under the curve (AUC), using an in-house script. The most convincing binding
modes, as well as an AUC = 0.776 were observed for model of the highest DOPE score,
which was consequently chosen for further studies, and denoted as model 1.

Template recognition and alignment correction

In order to find the most suitable protein template for A3AR receptor model, its sequence
obtained from UNIPROT [3] (sp_P33765) was used for a BLAST search using two online tools:
SwissModel [4] and ProteinBlast (NCBI) [5]. In both cases, default search modes for the most
similar PDB crystal structures was used. After comparison of the results, three templates were
chosen: 3EML (2.60 , 39,86% identity), 2YDV (2.60 , 42,6% identity) and 3VG9 (2.70 ,
43,34% identity). Template proteins were chosen according to their highest crystallographic
resolution, leading to reliable models with quite high confidence.

Iteration
Despite its high enrichment and acceptable binding modes of active set of ligands, possible
steric clashes between ligands and amino acids were observed, as well as narrowing of the
bottom part of binding pocket. Therefore in first instance docked ligands were minimized
using SZYBKI module (OEDocking) and the homology model of the protein was minimized
(with ligand) using CHARMM. As several high energy poses and similar docking behavior
was observed for the set of tested ligands, another modeling approach was then
undertaken.
Due to the fact that ZMA ligand appears to be inactive towards A3AR and therefore might
influence the shape of binding pocket during modeling, active ligand ZINC 12533962 (A 3AR
Ki = 40 nM) was placed manually into the crystal structure of 3EML instead. As an active
conformation of the ligand, the previously obtained pose from docking to A1R [2] was
used. Such prepared protein served as a template using the same alignment as for model
1, excluding 2YDV and 3VH9 sequences and including the ligand and its position while
modeling.

Figure 1. Structure alignment of 2YDV, 3VG9, 3EML and the sequence of human adenosine A3 receptor (AA3R).

Backbone generation
Protein structures were pre-processed using the PyMOL software [6]: ligands, co-crystallization
agents, lysozyme (3EML), and the allosteric antibody (2YDV) were removed. Thus obtained
protein sequences were aligned using the PROMALS 3D [7] online tool. (Figure 1).
The resulting alignment, after visual inspection (position of transmembrane domains, possible
disulfide bridges), and small correction was used as an input for MODELLER [8].
As an outcome, ten models were constructed using the above-described input alignment.
Models were characterized by relatively high DOPE (Discrete Optimized Protein Energy) and
GA341 (reliability of a model, derived from statistical potentials) scores scoring functions
implemented in MODELLER. Three best DOPE scoring models (GA341 score = 1.0 for all
obtained models) were used for further studies.

Side chain modeling

Each of the models was then aligned to
3EML crystal structure (Figure 2) and
inspected carefully visually using UCSF
Chimera software [9]: an orientation of the
side chain of ASN2506.55 and other binding
pocket amino acids and their possible,
acceptable rotamers (according to Dunbrack
library [10]), as well as by means of possible
mutagenesis data. On the other hand, also
inspecting transmembrane domains to
avoid: gaps, obvious steric clashes,
unnatural side chain amino acid folding, as
well as a preservation of the disulfide bonds
between CYS833.25 CYS 166.

Figure 3 The four obtained A3AR homology models. Binding pocket residues were depicted as thick sticks (only labeled in model 1
panel). The X-ray crystallographic structure of the template A2AR (3EML), is shown in dark green.

Figure 2. Structural alignment of three obtained homology proteins (light blue,

blue and magenta) and 3EML (green).

From the output of ten models, the best scoring one (model 2, molpdf: 1823.76904, DOPE
score -41914.38281) was chosen for evaluation. Due to possible steric clashes between
the ligand and TRP2436.48 in the bottom part as well as PHE168(EL2) in the top of the
binding pocket in freshly modeled homology receptor, again, the position
of ZINC 12533962 in 3EML structure was corrected manually and thus prepared receptor
served as a template for Modeller. The resulting model (model 3, molpdf: 3149.86035,
DOPE score: -41785.51562) showed also possible steric clashes between the ligand and
PHE168, thus its position in the template protein was again corrected, and the protein was
again remodeled.
To the output model (model 4, molpdf: 1593.80505, DOPE score: -41855.65625) after
visual inspection, the set of actives and inactives was docked (using the same parameters
as for model 1), resulting in AUC = 0.844 . This model (4), without further refinement and
minimization, was chosen for upcoming docking studies (data not shown). Binding pockets
of all four obtained homology models are presented in Figure 3.

Acknowledgement
Authors acknowledge the support of COST Action: CM1207 - GLISTEN. Presented herein studies were financed on a basis of
STSM, grant No ECOST-STSM-CM1207-020215-054712
[1] Popoli, P.; Pepponi, R.; CNS Neurol. Disord. Drug Targets, 2012, 11(6), 664-674 [2] Kolb, P.; Phan, K.; Gao, Z.-G. et al.; PLOS one. 7, 2012, 11, e49910. [3] http://www.uniprot.org/ [4] http://swissmodel.expasy.org/ [5] http://blast.ncbi.nlm.nih.gov/Blast.cgi/ [6] The PyMOL Molecular Graphics System, Version 1.5.0.4 Schrdinger, LLC.
[7] Pei, J; Bong-Hyun, K; Grishin, N.V.;. Nucleic Acids Res. 2008, 36(7), 22952300 [8] Sali, A.; Blundell, T.L.; J. Mol. Biol. 1993, 234, 523-531. [9] Pettersen, E.F.; Goddard, T.D.; Huang, C.C. et al.; J Comput Chem. 2004, 25(13), 1605-12. [10] Dunbrack Jr., R.L.; Curr. Opin. Struct. Biol. 2002, 12, 431-440. [11] Bento, A.P.; Gaulton,
A.; Hersey, A. et al.; Nucleic Acids Res., 2014, 42, 1083-1090. [12] OEDocking v3.0.1, OpenEye Scientific Software, Santa Fe, NM. http://www.eyesopen.com/; McGann, M.; J. Comput. Aided Mol. Des., 2012, 26, 897-906.; SZYBKI 1.9.0.3: OpenEye Scientific Software, Santa Fe, NM. http://www.eyesopen.com