Critical Analysis of the Successes and Failures of Homology Models of G-protein coupled receptors

Abstract

We present a critical assessment of the performance of our homology model refinement method for G-protein coupled receptors (GPCRs), called LITICon, that led to top ranking structures in a recent structure prediction assessment GPCRDOCK2010. GPCRs form the largest class of drug targets for which only a few crystal structures are currently available. Therefore accurate homology models are essential for drug design in these receptors. We submitted five models each for human chemokine CXCR4 (bound to small molecule IT1t and peptide CVX15) and dopamine D3DR (bound to small molecule eticlopride) before the crystal structures were published. Our models in both CXCR4/IT1t and D3/eticlopride assessments were ranked first and second respectively by ligand RMSD to the crystal structures. For both receptors, we developed two types of protein models: homology models based on known GPCR crystal structures, and ab initio models based on the prediction method MembStruk. The homology based models compared better to the crystal structures than the ab initio models. However a robust refinement procedure for obtaining high accuracy structures is needed. We demonstrate that optimization of the helical tilt, rotation and translation are vital for GPCR homology model refinement. As a proof of concept, our in-house refinement program LITiCon captured the distinct orientation of TM2 in CXCR4, which differs from that of adrenoreceptors. These findings would be critical for refining GPCR homology models in future.

INTRODUCTION

G-Protein Coupled Receptors (GPCRs) are membrane bound proteins that transduce signal from outside to inside the cell thus leading to a variety of biological processes. They form the largest super family of drug targets and therefore their three-dimensional structural and dynamic information are of great interest to researchers and the pharmaceutical industry to design effective drugs targeting GPCRs. In the past five years there have been several breakthroughs in solving GPCR structures both in the form of inactive and active conformations of the receptors^1-6. However the reported crystal structures represent a small fraction of the total number of GPCR species in the mammalian genome. As the number of available crystal structures increases, homology or comparative modeling methods can be used to derive structural models of other GPCRs. However the accuracy of homology models needs to be improved to the level of accuracy of high resolution crystal structures before using the models for drug design. There is a dire need for computational methods that can “refine” the homology models to capture the nuances of high resolution structures. In this context, prior to the publication of the crystal structures of the human chemokine receptor CXCR4 and dopamine receptor D3DR, a structure prediction assessment named GPCRDOCK2010 was announced⁷. The objective of this contest was to assess the status of current computational methods for predicting the structures of GPCRs and especially the ligand binding sites. As part of this assessment we submitted fifteen receptor models: five CXCR4 models with a small molecule antagonist IT1t bound, five CXCR4 models with a peptide CVX15 bound, and five models of the dopamine D3DR with the antagonist eticlopride bound. For small molecule bound D3DR and CXCR4, our predicted top ranking receptor-ligand complexes showed close agreement with the crystal structures in terms of ligand binding pocket and ligand pose prediction. Although the predicted binding pose for the peptide CVX15 was ranked second among the contestants, its agreement with the crystal structure was substantially less compared to the small molecule.

Three out of five top ranking receptor models submitted for CXCR4 and D3DR were modeled using the homology modeling package MODELLER⁸. The homology models were refined subsequently using our in-house homology model refinement method LITiCon^9,1. While D3DR dopamine receptor is close in sequence identity to the β₁ and β₂-adrenergic receptors (denoted as β₁-AR and β₂-AR hereafter), CXCR4 did not have a closely related homologous template that can be used for homology modeling.

In this paper we demonstrate that optimization of the helical tilt, rotation, translation and gyration of the seven transmembrane (TM) helices lead to substantial refinement of the GPCR homology models even with remote homology templates. We present a critical analysis of our predicted receptor models and the binding site of the antagonists as well as provide insights into how the modeling can be done for GPCRs that have little homology to known structures. During this work, we also developed a more detailed procedure “GPCRCompare” for comparing two GPCR structures. This procedure compared the relative TM helical tilt, rotation, gyration and translation of all the seven TM helices. These procedures described in this paper would be extremely useful for drug design in academia and pharmaceutical industry.

METHODS

GPCR structural comparison

To compare the helical orientations of two GPCR structures, the two structures are aligned using the Cα coordinates of the TM regions. Since the loops are more structurally divergent than the TMs, the loops are not included in the alignment. The relative orientation of each TM helix with respect to another is composed of three independent components: translation of the helix, tilt about the center of mass and rotation about the helical axis. We calculate these individual components for each TM in one receptor structure relative to the corresponding TM in the other one (Fig. S5).

CXCR4 Structural Models

Of the top five structural models of CXCR4 that we submitted to GPCRDOCK 2010 Assessment, three were homology based models, and two were derived using an ab initio method called MembStruk¹¹. The details of the MembStruk method are discussed elsewhere¹¹. In brief, the MembStruk method starts with a template of the 7 transmembrane domains based on β₂-AR crystal structure. It then optimizes the packing of the TM domains by refining the rotational orientation of each TM helix over ±180°. The top scoring energy minima for each TM helix are combined to identify the best scoring combination of rotation angles.

The three homology based models were derived using MODELLER⁸ and with three different sets of templates. One of the models was built using both A_2A adenosine (PDB ID: 3EML) and β₂ adrenergic (PDB ID: 2RH1) receptors as templates^4,2 , and the remaining two models were built using β₁-AR (PDB ID: 2VT4) and β₂-AR as templates respectively^5,2. In the β₁-AR based model, TM2 and TM4 were optimized using the LITiCon method. TM2 was rotated in increments of 5° from −40° to 40°. At each rotated conformation the side chains were optimized using SCWRL3.0¹² and the total energy of the receptor was minimized using conjugate gradient minimization. The Van der Waals (vdW) energy showed a minimum at a rotation angle of −15° as shown in Figure 5a. Hence we selected this conformation for TM2. A similar procedure was followed for TM4 as well. For TM4, the vdW energy was flat near the initial structure, while the potential energy showed minima at −10° and 35° as shown in Figure 5c. We selected the TM4 conformation corresponding to a 35° rotation, which was the global minimum in energy for TM4 rotations. We docked the IT1t ligand to all the homology models with different templates and selected the β₁-AR based homology model for submission, because it gave the best ligand docked poses satisfying the criteria used for selecting the best ligand pose as detailed in the ligand docking section below. The best structure was then minimized using the minimizer in NAMD and CHARMM force field¹³. The major difference between the MembStruk and LITiCon procedure for rotational optimization is that LITiCon performs simultaneous optimization of rotational orientations of all the seven helices, while the MembStruk method performs this optimization serially for each helix.

Figure 5.

Optimization of TM2 and TM4 orientations of CXCR4 using LITiCon conformational sampling; (a & c) plots of potential energy of TM2 and TM4 as functions of helical rotation; the minima in energy are highlighted. The original and optimized conformations are marked using same color code as Fig. b & d. (b & d) Overlay of original homology model with the optimized structure and the crystal structure of CXCR4 with IT1t bound.

D3DR Structural Models

Of the five D3DR structural models, the first three models (by ranking) were derived through homology methods using β₂-AR as template². Since the ECII loop of D3DR shows little homology to β₂-AR, this loop was modeled using the ab-initio loop modeling procedure in MODELLER⁸ without using any structural template and an imposed disulfide constraint between C103 and C181 (EC-II). Each of the structures was minimized using the DREIDING force field¹⁴, before subjecting to the docking protocol. Rigid body optimization was not performed in the case of D3DR because of the high homology with β₂-AR.

Docking of small molecule ligands

The small molecules IT1t for CXCR4, and eticlopride for D3DR were built and optimized using LigPrep module in Maestro. Several ligand conformations were generated outside the protein and docked using Glide XP (Schrödinger Inc). 10 docked conformations were retained for each ligand conformation. The docked conformations were clustered using Xcluster (Schrödinger Inc.) with a cluster distance cutoff of 1.5A. The criteria to select the poses were based on the mutation data available for cyclam and non-cyclam compounds in CXCR4^16-18. For selecting the eticlopride docked poses, the mutation data on D2DR¹⁹ (eticlopride binding not affected by any of the serine mutations on TM5) was considered.

RESULTS AND DISCUSSION

Comparison of the predicted binding site of the small molecule IT1t in CXCR4 and eticlopride in D3DR to their crystal structures

Figures 1a and b show the CRMSDs of the predicted ligand poses (eticlopride and IT1t) to those in the crystal structures. The ligand CRMSDs were calculated by aligning the predicted receptor models to the corresponding crystal structures. This gives the ligand CRMSD in place, which indicates how close the predicted ligand pose is to the crystal structure pose. We also calculated the ligand CRMSDs by aligning the predicted ligand structure to the ligand pose from the crystal structure (superposition with alignment). This gives the accuracy of the internal structures of the predicted ligand poses. Both IT1t in CXCR4 and eticlopride in D3DR show close agreement with the placement of the ligand in the receptor (CRMSDs for ligand heavy atoms: 1.6Å for eticlopride and 2.7Å for IT1t). For all the comparison to the crystal structures we used the PDB ID: 3ODU for CXCR4 and 3PBL for D3DR. The predicted ligand conformations show internal RMSD of 0.6Å for eticlopride and 1.8Å for IT-1t, as shown by alignment of the ligands to the crystal poses. Fig. 1c and d show the orientations of the ligands within their respective binding cavities as compared to their crystal structures. Our predicted ligand binding pose captures 50% and 90% of all the protein ligand contacts within 5Å of the ligand in the CXCR4/IT1t and D3DR/eticlopride crystal structures respectively. Although, we missed about 50% of the binding site residues in case of CXCR4/IT1t, many of these missed residues are in the EC loops. Among the residues, in the TM regions, our prediction rate was much higher. We predicted 70% of the TM residues that are within 5Å of IT1t. Many of the loop residues were not identified due to the lack of accuracy in our loop predictions. However most of the key residues that govern the binding of small molecule antagonists (such as IT1t) in CXCR4 (as observed from mutation studies) are located in the transmembrane domains. This in conjunction with the challenges associated with the CXCR4 target indicate that the level of accuracy in the predicted models is well suited for predicting site directed mutagenesis candidates as well as for studying the Structure-Activity-Relationships (SAR). The number of residues was calculated using PyMol²⁰, unlike in the GPCRDOCK2010 assessment where a soft cut-off was used⁷.

Figure 1.

Comparison between the predicted and crystal structure ligand poses superposed both in place (individual receptor structures aligned) and structurally aligned; (a) eticlopride; (b) IT1t; (c) & (d): Comparison between the predicted and crystal structure binding sites of (c) eticlopride in D3DR; (d) IT1t in CXCR4; Residues within 5Å of the ligands are displayed.

Detailed Predictions of the Binding Site residues

Table I shows the number of residues within a certain cutoff distance from the ligand in CXCR4 and D3DR predicted structural models. It is clear that about 50% of the protein ligands are captured in the CXCR4/IT1t structure and about 90% of the contacts are captured in D3DR/eticlopride. There are two possible protonatable nitrogens in the compound IT1t. Our model captures the tight contact of one of these amine groups with the D97^2.63 (2.7Å) and the weaker contact with E288^7.39 as seen in the crystal structure. In the predicted ligand pose, the imidothiazole group is positioned in the same location within the binding site as the crystal structure but rotated leading to a larger CRMSD for this part of the ligand.

Table I.

Percentage of receptor ligand interactions captured in our predicted binding poses.

Receptor		CXCR4	D3DR
Residues within 3Å	No. of residues in crystal structure	2	1
	% captured in homology model	50%	100%
Residues within 4Å	No. of residues in crystal structure	12	14
	% captured in homology model	42%	86%
Residues within 5Å	No. of residues in crystal structure	16	20
	% captured in homology model	50%	90%
Residues within 6Å	No. of residues in crystal structure	22	27
	% captured in homology model	59%	85%

In order to evaluate the accuracy of predicting the binding sites, we calculated the RMSDs of the amino acids within 5Å of the ligand in the predicted receptor structures to those in the crystal structures. The RMSDs of the predicted models of CXCR4 and D3DR are shown in Table II. In the CXCR4 predicted model, with the exception of W94^2.60, all the residues in the TM helices show low CRMSDs to the crystal structure (<3Å), while the loop residues show larger deviations. From Fig. 1d, the backbone of W94^2.60 is situated close to that of the crystal structure. However the side chain of W94^2.60 shows a large deviation from that of the crystal structure, which contributes to the large RMSD of this residue. In D3DR, all binding site residues are part of TM domains. All of these residues show low CRMSDs to the crystal structure thus highlighting the accuracy of our prediction.

Table II.

CRMSD of residues in the binding site for CXCR4/IT1t and D3DR/eticlopride.

CXCR4/Residue	CRMSD Å	D3DR/Residue	CRMSD in Å
W942.60	5.5	D1103.32	1.2
D972.63	2.9	S1925.42	2.2
V1123.28	2.5	S1935.43	1.3
Y1163.32	1.6	S1965.46	1.7
E2887.39	2.3	W3426.48	1.3
W102ECL1	6.3	F3456.51	1.1
R183ECL2	10.9	F3466.52	1.1
I185ECL2	4.1	H3496.55	2.9
C186ECL2	5.6	Y3737.43	1.1
D187ECL2	8.2

Pitfalls of docking small ligands to GPCRs

Clearly between the two receptor targets, prediction of ligand binding to CXCR4 was more challenging, since there were no prior site directed mutagenesis studies on the IT1t or related ligands. The binding pocket of CXCR4 consists of four charged residues (D97^2.63, D171^4.60, D262^6.58, E288^7.39), all of which show strong reduction of antagonist binding upon mutation to neutral residues^16,18. Docking of IT1t showed two possible energetically favorable binding sites. One located in the region between D171^4.60, D262^6.58, E288^7.39 and the other located among the residues D97^2.63, D262^6.58, E288^7.39. The literature on the effect of site directed mutagenesis on binding of cyclam based compounds leads to ambiguous positioning of the IT1t ligand^16-18. Thus determining the principal salt bridge for IT1t using preexisting knowledge from other small molecules is non-trivial. Based on our previous knowledge of antagonist binding in chemokine receptors^21-24, we hypothesized that the nitrogen atom on the imidothiazole group (one on the same side as the sulfur atom) could form a hydrogen bond (HB) with D97^3.63 and E288^7.39. Since D97^2.63 was important in non-cyclam antagonist binding in CXCR4¹⁸, and E288^7.39 was important for antagonist binding in many chemokine receptors²¹, we selected those ligand poses which satisfied the criteria of contacting D97^2.63 and E288^7.39 and obtained 48 ligand poses. We selected the final ligand pose by visualizing each of the 48 poses and using criteria such as how well the ligand fits into the binding cavity, and whether there is any unusual internal strain in the ligand pose. In contrast, predicting the eticlopride pose in D3DR involved clustering the docked poses into distinct clusters based on CRMSD and then selecting the top pose by binding energy from each cluster. However the final ligand poses were selected by visualization in conjunction with the binding energy.

Based on our docking predictions, we found that while docking methods were efficient in sampling accurate ligand poses (ones close to the crystal structure poses), binding energy of the docked poses were not accurate enough to score the best structure at the top. The inaccuracies in the binding energy calculation could be contributed by the structural defects of the homology models as well as our approximate method of estimating energy, which involves limited minimization and crude approximation of desolvation. At the end, experimental data from binding site mutations were useful in identifying the correct ligand pose.

Analysis of receptor models and crystal structures

The CRMSD in coordinates of the prescribed set of atoms when comparing two GPCR structures gives an averaged quantitative comparison while missing the differences in the structural features such as the seven helical rotations, tilt, translation and gyration that are important for the function of the GPCR. Therefore, we derived a geometrical method, “GPCRCompare” to compare two GPCR structures in terms of these helical properties. Each of the seven helical TM regions will be compared in terms of the helical rotation, tilt and translation. We have analyzed the merits and pitfalls of our submitted CXCR4 and D3DR models by comparing them to the crystal structures. We have also compared the crystal structures of D3DR and CXCR4 to the crystal structure of β₂-AR (PDB ID: 2RH1) in terms of the seven helical properties. More importantly, we have demonstrated how optimization of these helical translation rotation, tilt and gyration leads to refinement of the homology models for two GPCRs with low sequence homology.

TM Helical Rotation and Tilts capture most of the differences between GPCR crystal structures

Analysis of the structural differences in the seven TM helices for the CXCR4 and D3DR crystal structures with respect to β₂-AR in the inactive state are detailed in the next section. Summarizing these results, it is seen that the maximum difference in structures of CXCR4 and β₂-AR stems from helical rotation, and tilt of TM helices 5 and 6. To demonstrate this point clearly we took the TM5 and TM6 of CXCR4 crystal structure and performed the helical translation, rotation and tilts differences between CXCR4 and β₂-AR crystal structures in TM5 and TM6 and got structures that are close to β₂-AR. Figure 2 demonstrates the result of these operations on the receptor models. It is seen that after successive rigid body transformations TM5 and TM6 of CXCR4 show substantial improvement resulting in less than 0.5Å CRMSD to β₂-AR at the end of these helical movements. Thus we find that the helical tilts and rotations contribute substantially to the structural diversities of TM5 and 6 and therefore optimization of helical rotations, tilt and translation would lead to substantial refinement in GPCR homology models.

Figure 2.

Improvement in CRMSD of TM helices 5 and 6 of CXCR4, to those of β₂-AR after successive rigid body transformations (a) TM5; (b) TM6.

Comparison of all the predicted models to crystal structures and pitfalls of the structure prediction methods

For the GPCRDOCK2010 challenge, we used both homology modeling technique⁸ and the ab initio GPCR structure prediction method MembStruk¹¹ to generate the receptor structural models. Among the five predicted models for CXCR4, models 1, 2 and 5 (by our ranking) were homology models, and structures 3 and 4 were predicted using MembStruk. Among D3DR models, structures 1-3 were homology models, and structures 4 and 5 were using MembStruk. The CXCR4 homology model was further subjected to rigid body optimization of TM2 and TM4 using the software LITiCon¹. No LITiCon optimization was performed with the D3DR models because D3DR has a sufficiently high homology with the template β₂-AR. We have compared all the receptor structural models to their respective crystal structures. Table S2 summarizes the major differences in the orientations of TM helices of the CXCR4 and D3DR models to the respective crystal structures. Both the homology and MembStruk CXCR4 models show large RMS deviations from the crystal structures in the TM regions. However the MembStruk model shows a larger RMSD (6Å) compared to the homology model (4.9Å). Comparing the RMSDs of individual TM domains of the homology model, TM1 shows the lowest deviation with the crystal structure (3.4Å), while TM5 and TM6 show the worst deviations (6Å). The RMS deviation of TM5 is due to a misalignment of the sequence in the intracellular region, and most of the RMSD in TM6 in the homology model comes from the intracellular domain. TMs 1, 4, 5 and 7 in the CXCR4 homology model show large RMSDs when the TMs are individually aligned to the corresponding TMs in the crystal structure. For these TMs, the differences with the crystal structure are contributed by both packing and internal deformations in the helices. In contrast, the structural deviations of TMs 2, 3 and 6 are mostly due to packing differences, since their internal structures are close to those in the crystal structures.

In the MembStruk model, all TM regions show larger RMS deviation with the crystal structure compared to the homology model, with the largest deviations being observed in TM2 and TM6. Both these TMs show large internal deformations (high aligned RMSDs in table S2). Hence their structural deviations are contributed by helical kinks as well as packing differences. The optimization of the helical degrees of freedom in the MembStruk models did not yield rotation angles close to the crystal structure, due to the inadequacy of the energy functions used in structure prediction or refinement to identify the best energy structure as the one close to the crystal structure. This has been observed previously for globular proteins (Critical Assessment of techniques for Structure Prediction http://predictioncenter.org/) and hence more comprehensive energy function that take into account membrane solvation are required. Moreover, the internal structures of TMs 1, 2, 4, 6 and 7 are different than the ones in the crystal structure (as reflected by their high aligned RMSDs), which also contribute to the structural differences with the crystal conformation.

Among the D3DR models, both homology and MembStruk models are closer in RMSD to the D3DR crystal structure (1.5Å and 2.7Å respectively) in comparison to the CXCR4 models. In the homology model, TM4 shows a 16° rotation relative to the crystal structure. In the MembStruk model, the largest deviations are shown by TM1, TM2 and TM4 (3-3.2Å). The RMSD in TM1 is contributed by a 2.3Å translation and a 45° rotation. TM2 shows high internal deformation, while TM4 shows a 2.5Å translation. For the remainder of this section, we will discuss the homology models of the two receptors. Thus we observe that the homology model with subsequent rigid body optimization would produce better results than ab initio methods.

Fig. 3 shows the CRMSD for each residue of CXCR4 and D3DR from their respective crystal structures. In both models, the TM regions show much closer agreement with the crystal structures compared to the loop regions. In the CXCR4 model, the extracellular loop 2 (ECL2) proximal to TM4 shows higher CRMSD (~15Å), while the part near TM5 shows a relatively lower CRMSD. Overall, our loop predictions place the EC loops covering the entire extracellular surface of the TMs. In the crystal structure, the ECL2 is located closer to TMs 2 and 3 while ECL3 is located outwards between TMs 6 and 7. This opens up a cavity in the EC surface of the crystal structure with direct access to the ligand binding pocket.

Figure 3.

CRMSD of the C_α atoms of each residue in our best models of CXCR4 and D3DR compared to their respective crystal structures. For CXCR4 we used the PDB ID: 3ODU and for D3DR PDB ID: 3PBL

Comparison of the location of the helical kinks and the kink angles in each TM calculated using the software PROKINK²⁵, are shown in Table S1. The kink angles are compared between the models and the corresponding crystal structures. Each of the TM regions in the CXCR4 crystal structure shows moderate to large kinks. TM7 shows the largest kink (45°) and TM3 shows the smallest kink (19°). The CXCR4 homology model shows good overall agreement in terms of kink prediction with the exception of TM1. TM1 in the crystal structure shows a 24° kink near G52^1.46. This kink is not observed in the homology model. Also TM6 in the homology model shows a smaller kink (23°) compared to the one in the crystal structure (40°). In the D3DR model, all TMs except TM1 and 3 show significant kinks. TM2 shows the largest kink near V82^2.57. For D3DR all kinks observed in the crystal structure were correctly predicted in our best model (Table S1).

The helical tilt, rotation and translation are shown schematically in Fig. S5 of the Supplementary material. Fig. S1 and S2 of the Supplementary material show the detailed structural comparison of the predicted models of CXCR4 and D3DR with their respective crystal structures in terms of helical rotation, tilt, and translation for each TM helix. The signs of the rotation and tilt angles in Fig. S1, S2 and S3 represent the direction of movement relative to the protein core while viewed from the extracellular (EC) side. A positive tilt indicates tilting towards the protein core (termed inward henceforth), whereas negative tilt shows motion away from the protein core (outward). Likewise, positive rotation represents anticlockwise motion and vice versa. Fig. 4 shows a schematic representation of the major differences between the TM helices of the crystal structures. Among the TM helices of CXCR4 model, TM5 shows the largest deviation from its crystal structure, with a 3.5Å translation, 17° tilt and a 58° rotation. The structural differences of TM6 with crystal structure comes mainly from helical rigid body deviations, whereas for TM4, 5, and 7, these differences stem from both changes in the rigid body degrees of freedom and internal helical deformations. For the D3DR model, the TM helices show closer agreement with the crystal structure. All the helices show negligible translation (<1Å) and tilt (<5°), except TM1.

Figure 4.

Comparison of rigid body structural differences (translation, tilt and axial rotation) between the homology models and the corresponding crystal structures; (a) CXCR4; (b) D3DR; Helices are represented by cylinders; block arrows represent center of mass translations, curved arrows represent tilts and rotations. Figures are schematic, not to actual scale.

Merits and current pitfalls of LITiCon, the helix rigid body optimization procedure for GPCR homology models

At present, homology modeling is the most widely used technique for predicting the structures of GPCRs. However modeling GPCRs using homology to known crystal structures poses a unique challenge, since not many high-resolution crystal structures are available for distinct receptor classes. We demonstrate how the LITiCon method optimizes the helical rotations of the predicted models of CXCR4 compared to its crystal structure in the GPCRDOCK2010 assessment.

In CXCR4, the TM helices TM2 and TM4 are shown to be important for binding non-peptide antagonists from mutation studies^16-18. Particularly W94^2.60 and D97^2.63 on TM2 and D171^4.60 on TM4 have been implicated in the binding of non-cyclam antagonist¹⁸. In the CXCR4 homology model based on β₂-AR template, the TXP motif on the EC side of TM2 shows a different kink compared to the β₂-AR crystal structure, which places W94^2.60 facing the membrane instead of the binding pocket in the homology structure of CXCR4. We used the LITiCon method¹ to optimize the rotational orientation of TM2. Fig. 5a shows the potential energy of the CXCR4 model for different rotational orientations of TM2. A rotation of 20° relative to the original state was found to be the minimum in energy and was selected as the refined model. A comparison of the unrefined homology model, the refined TM2 model and the TM2 structure from the CXCR4 crystal structure shown in Fig. 5b reveals the improvement that is achieved using the rigid body optimization in LITiCon. The TM2 rotational orientation of the predicted model is 3° different in the axial rotation compared to the crystal structure. These subtle structural improvements are important for optimizing the binding pocket as seen in our predicted binding pocket. Our predicted ligand binding pocket was the closest to the crystal structures among all the models submitted to GPCRDOCK2010.

LITiCon optimization was also performed on TM4 of the CXCR4 homology model. The purpose of this was to optimize the position of D171^4.60 that was facing outward in the original homology model (Fig. 5d). Fig. 5c shows the plot of the potential energy of TM4 as function of axial rotations and Fig. 5d shows the comparison of the original and refined TM4 conformations to the crystal structure. The potential energy landscape shows two distinct minima and we selected the global minimum (rotation angle of 40°) as the refined conformation. In the refined homology model, D171^4.60 has moved inward into the binding cavity thus improving the agreement with the crystal structure. It should be noted that for these predicted homology models we performed rigid body optimization for some of the helices and not all the seven TM helices. However, before we move towards performing optimization of rigid body orientations for all the seven helices, one needs to develop a robust scoring function that identifies native like structure as low energy conformations.

However due to the difference in helical kinks, the internal structure of TM4 is different in the homology model compared to the crystal structure, as shown in Fig. 5d. Therefore rigid body optimization would be effective provided we are able to predict the internal structures of the TM helices accurately. Since homology models mimic the kinks and deformations of the template structures rather than the target, efficient computational protocols are needed to optimize the internal deformations of the TM helices such as kinks and bulges²⁶. Kink optimization combined with rigid body refinement techniques like LITiCon could lead to a robust platform for homology model refinement.

Comparison of the crystal structures of CXCR4 and D3DR to the crystal structure of β₂-adrenergic receptor in terms of rigid body movements

Table III shows the RMSD of the C^α atoms (CRMSD) of the seven helices of D3DR (PDB ID: 3PBL) and CXCR4 crystal structures (PDB ID: 3ODU) from the β₂-AR crystal structure (PDB ID: 2RH1). For each receptor, the first column shows the CRMSD for each helix when the entire receptor structures are aligned (called “in place” in Table I), and the second column shows the CRMSD when each helix is individually aligned to the corresponding helix in β₂-AR. The “in place” CRMSD gives the overall difference in RMSD between the two receptors, while the “aligned” CRMSD shows whether the structural differences are due to internal deformations or rigid body orientations. The CXCR4 helices have a higher CRMSD to β₂-AR compared to those of D3DR as anticipated, since their sequences have less than 20% similarity. The TM helices 1, 4, 5 and 6 in CXCR4, show large CRMSD from β₂-AR (>2Å) with TM1 and 4 showing the highest CRMSD. TM5 and 6 of CXCR4 when aligned to those of β₂-AR (table I, “aligned”), show CRMSDs below 2Å implying that the structural differences in these helices arise mainly from rigid body orientations. In contrast, TM1 and TM4 do not show a sizable reduction in CRMSD on alignment with β₂-AR. Thus their structural differences arise from both rigid body orientations and internal deformations, namely kinks and bulges. Among the TM helices in D3DR, only TM4 and 6 show CRMSDs greater than 2Å. All D3DR helices show “aligned” CRMSDs less than 2Å (table III). Therefore these differences are caused by rigid body orientations and not by structural deformations.

Table III.

CRMSDs in Å of the Cα atoms of the TM helices of CXCR4 and D3DR with respect to β₂-AR crystal structure. CRMSD "in place" denotes the CRMSD of a TM helix when the TM domains of the entire receptor are aligned to those of □₂-AR. CRMSD "aligned" denotes the CRMSD of the C_α atoms of the TM helix aligned to the corresponding TM region of β₂-AR.

TM region	D3DR		CXCR4
	In place	aligned	In place
TM1	1.7	0.7	4.5
TM2	1.0	0.7	2.0
TM3	1.0	0.6	1.5
TM4	2.0	1.4	3.4
TM5	1.8	1.5	2.3
TM6	2.3	0.8	2.9
TM7	1.3	0.7	1.5

Further analysis of the rigid body movements such as helical tilt, rotation and translation of the TM helices of CXCR4 and D3DR with respect to β₂-AR characterizes the structural changes more accurately (Supplementary Fig. S4). The structural differences of TM1 in CXCR4 are in helical translation (1.4Å), an inward tilt (17°) and a small rotation (10°). TM2 shows a translation (0.9Å) and a rotation (20°), but no tilt. TM4 shows translation of 0.5Å and tilt (12°), with a large rotation (25°). The rigid body differences of TM5 are caused by a large rotation (30°) and relatively minor translation and tilt. TM6 shows mainly translation (0.9Å) and little tilt (9°), but no rotation. In D3DR, TM4 shows a small rotation (10°), while TM6 shows a small translation (0.6Å), tilt (5°) and rotation (8°). A schematic representation of these major structural changes is shown in supplementary Fig. S5.

CONCLUSIONS

Our predicted D3DR and CXCR4 structures with bound ligands were ranked at the top in the GPCRDOCK2010 assessment. Here we have shown that optimization of TM helical tilt, rotation, translation and gyration of the homology models using the LITiCon method captures the structural differences between the template structure and the homology model, leading to structure refinement for GPCRs. TM2 in chemokine receptors were predicted to have a distinctly different conformation from the biogenic amine receptors. Our predicted structural model has TM2 rotational orientation within 3° of the crystal structure.

The CRMSD is a simplistic measure of comparison of two GPCR structures that lacks a detailed comparison of the structural nuances critical to refinement of homology models. We have developed a geometrical method, GPCRCompare, to compare two GPCR structures in terms of helical tilt, translation, rotation and gyration degrees of freedom. We have demonstrated the use of this method by comparing crystal structures of D3DR and CXCR4 with β₂-AR. TM1, TM4, TM5 and TM6 were the most structurally diverse between CXCR4 and β₂-AR, while TM3 was the most conserved. TM4 also showed large internal deformations, whereas the structural differences of TM5 and TM6 were mainly from rigid body orientations and not from internal deformations. We have also demonstrated that optimizing the rigid body orientations (i.e. translation, tilt and rotation) of TM5 and TM6 of CXCR4 leads to improved CRMSD with the β₂-AR crystal structure. Hence rigid body optimization along with kink prediction could be useful in refining crude homology models. We have also discussed the structural comparison of our predicted homology models to the crystal structures and discussed the merits and failures of our prediction scheme.

The ligand docking methods are able to predict the ligand docked poses but the energy functions are not adequate to pick the correct pose. We had to use prior knowledge from other chemokine receptors in selecting the best ligand pose. In summary, refinement of homology models based on rigid body degrees of freedom is critical to deriving accurate homology models for GPCRs.