Since the alignment of compound structures plays an important role in developing successful 3D-QSAR models [39], two rules (both ligand-based and docking-based) were adopted to align the dataset to derive reliable models. The results obtained from both models using the same training set of 81 compounds are summarized in Table 1. A number of statistical parameters, i.e., the Q², non-cross-validated correlation coefficient (R²_ncv), SEE, and F-statistic values, are analyzed to evaluate the quality of the models.

In both CoMFA and CoMSIA analyses, ligand-based alignment modeling leads to models with larger R²_cv, R²_ncv, R²_pred values than the corresponding models obtained by the receptor-based alignment modeling. Therefore, we mainly focussed on the ligand-based 3D-QSAR models for further analysis. In addition, since the five (steric, electrostatic, hydrophobic, and H-bond donor and acceptor) field descriptors may not be completely independent of each other and such dependency among individual fields may reduce the model significance and generalization [40,41], all possible combinations of the descriptors were used to derive models for avoiding the risk of omitting possible optimal model and to explore the best combination use of the descriptors for model generation. Finally, a CoMSIA model established by using the steric, electrostatic, hydrophobic and hydrogen bond donor field descriptors appears to be superior to all other models derived, whose statistical results are listed in Table 1. Using seven PLS components, this model yields statistical results of Q² = 0.501, R²_ncv = 0.912, SEE = 0.250 and F = 108.309 with steric (12.4%), electrostatic (38.7%), hydrophobic (24.4%) and H-bond donor (24.5%) field contributions, proving its correct internal predictive capability.

Generally speaking, a Q² > 0.5 is considered proof of acceptable internal predictive ability [42]. What’s more, the high R²_ncv and F values along with the low SEE values should also be considered as the foundation of a reliable QSAR model [43]. However, due to chance correlation or structural redundancy, sometimes it is found that some models derived from the training set molecules with randomized activity possess high Q² values, but show unfavorable predictivity for prediction of unknown molecules [33,44]. Hence, the extensively accepted leave-one-out (LOO) cross-validated Q² is insufficient to assess the predictive power of the QSAR models [45]. In light of such risks, we validated the models by predicting the activity of an external test set composed of 22 NOP agonists. As a result, a predictive coefficient R²_pred of 0.818 was achieved verifying the good external predictive efficacy of the model (Table 1). Figure 1 illustrates the correlation plot of experimental versus predicted pKi values of the training (filled red square) and test (filled green triangle) sets based on the optimal CoMSIA model. Clearly, a good correlationship is observed from this figure since the predicted values are almost as accurate as the experimental activities for the whole dataset, and all points are rather uniformly distributed around the regression line, indicating no existence of systematic errors in the method. This good agreement between the predicted and experimental activity data proves the satisfactory predictive ability of the CoMSIA model.

The 3D-coefficient contour plots are beneficial to identify important regions where some changes in the interaction fields can affect the biological activity, and may also be of help to identify the possible interaction sites of the biochemical system. Thus presently, the optimal ligand-based CoMSIA model is selected for each conformation to construct the stdev*coeff contour maps to view the field effects on the target features due to its good internal and external predictive powers. The maps generated depict regions having scaled coefficients greater than 80% (favored) or less than 20% (disfavored). To aid in visualization, the most active compound 32 is shown as template molecule with the contour maps (Figure 2).

The CoMSIA steric contour plot for the most active compound 32 is displayed in Figure 2A, where the sterically favored regions are shown in green and disfavored regions in yellow, respectively. As seen from this picture, some green contours are mapped near position-1 of ring A, positions-11 and -12 of ring D and positions-19 and -20 of ring E, suggesting that bulkier groups are favored at these positions. The green contour around position-1 is well consistent with the higher potency of compound 53 with a bulkier substituent (CH₃OC(O)CH₂-) (pK_i = 7.71) than compound 15 without substituent (pK_i = 7.63) at position-1 of ring A. The higher potency of compound 38 (pK_i = 8.85, with a CH₃ group) than 36 (pK_i = 7.96, without substituent) is also such a case. A few residues located around positions-11, -12, -19 and -20 lead to a large empty space at these positions (as shown in Section 2.3), which can interpret the presence of green contours at these positions. Two yellow contour maps appeared above ring C and between ring C and ring D, respectively, implying that bulkier substituents at these positions may decrease the activity; the reason may be that there is no such bulky substituent at these positions of both rings C and D in all molecules of the dataset.

Figure 2B depicts the electrostatic contour maps obtained from the CoMSIA model, where blue contours represent the favorable electropositive regions and red contours account for the favorable electronegative regions, respectively. A large blue contour extending from position-1 to position-4 indicates that electropositive groups are preferred here. Compound 62 with substituent of (CH₃)₂N(CH₂)₂- at position-1 shows higher activity than compound 56 with substituent of NH₂(CH₂)₂- due to the stronger electronegativity of the former substituent. Besides the large blue contour, a red contour can be seen near position-1 of ring A indicating that this position is sensitive to electrostatic substituents. This phenomenon may have something to do with the atom N (electronegative atom) of ligand 32 at position-1. In addition, the atom O (electronegative atom) of ligand 32 at position-5, may be the reason for the small red contour showing around here. The function and location (next to position-1) of atom O at position-5 may lead to the red contour appearing around position-1. Also a blue contour located at positions-11 of ring D, suggests the possible help of electropositive substituents, improving the binding affinity. Actually, this may be due to the existence of the electro-negative amino acids Asp130 and Asp209 in the binding pocket, as discussed later in Section 2.3, because of the lack of any substituents in these areas. In addition, a large red contour map is seen around position-8, which is caused by the electronegative atom N and an electro-positive amino acid Arg302 (as discussed later in Section 2.3).

Figure 2C shows the CoMSIA hydrophobic contour map, where the yellow (hydrophobic favorable) and white (hydrophobic unfavorable) contours represent 80% and 20% contributions respectively. Substitutions by hydrophobic groups like -Cl and -F at positions-18 and -19 are extended to the yellow contours resulting in a higher NOP activity, which can be illustrated by the example that compound 33 with -F at position-18, and compound 32 with -F at position-19 all exhibit higher activities than compound 29 without any substituent at either positions-18 or -19, respectively. The existence of atom F at position-19 may be the reason for the yellow contour extending. Furthermore, two large yellow regions are observed above positions-6 and -7 and around position-9 of ring C, respectively. Also some white regions are observed close to position-1 of ring A, ring B, positions-11 and -12 of ring C and position-17 of ring E indicating that hydrophilic groups here are helpful for the activity. The fact that compound 83 with hydrophilic group -(CH₂)₂OH has higher potency than compound 103 with hydrophobic group -NHBu at position-1 verifies this conclusion. What’s more, compound 54 with hydrophilic substituent-(CH₂)₂OH at position-1 shows higher binding affinity than both compounds 51 (with hydrophobic group -Bu) and 52 (with hydrophobic group i-Amyl-). The white contour around positions-11 and -12 of ring C and position-17 of ring E may be attributed to the existence of the hydrophilic residues (Asp110, Asp209, Asp130, Thr103, Thr305, Tyr309, Gln107 and Arg302) in binding pocket, as discussed later in Section 2.3 and shown in Figure 3B.

In Figure 2D, the CoMSIA hydrogen-bond donor plot, the cyan contours indicate regions where hydrogen bond donor substituents on the ligands are favored and the purple contours represent areas where hydrogen bond donor substituents on agonists are disfavored. As seen from the picture, a large cyan contour appears around position-1. Its appearance was due to the atom N at position-1 acting as an H-bond donor and interacts with the key amino residues around the position (as discussed later in Section 2.3). The only structural difference of compounds 72~79, lies in the substituent at position-1 (such as the EtNHCH₂CH₂- and BuNHCH₂CH₂- groups) which can serve as the H-bond donor for interaction with the surrounding environment. Thus, they all exhibit higher pK_i values than compound 80 with H-bond accepter substituent (

) at position-1, providing powerful proof for the conclusion. The atom N and atom O located at position-1 and position-5 respectively, both of which can act as H-bond donors, may be the reason for the large purple area around these positions.

Due to the important role in the rational design of a drug, docking is often used to find the optimal orientation of a ligand in the binding to its pharmaceutical target [46]. In our work, the whole dataset of 103 compounds were docked into the possible active site of NOP receptor crystal structures, and the optimal conformations of these compounds were determined. The results show that all molecules in the series were well placed in the active site demonstrating the rationality and reliability of the docking model. The binding mode of the highly potent compound 32 docked into the receptors is taken as an example and shown in Figure 3. As observed from this figure, the ligand core is anchored in the binding site via hydrophobic interactions and two H-bonds are identified as potential factors influencing the high binding affinity of compound 32. The specific binding interactions are analyzed in detail as follows.

As seen from the picture, compound 32 is actually docked into a basically hydrophobic pocket which is formed by Trp116, Trp211, Trp276, Val126, Pro207, Phe106, Phe220, Ile127, and Phe106. This finding reaches good agreement with the previous CoMSIA hydrophobic contour maps analysis in Figure 2C, that most hydrophobic residues around the binding ligand are consistent with those yellow contours appearing around position-9 of ring C and positions-18 and -19 of ring E. In addition, the presence of several hydrophilic residues (Asp209, Asp130, Thr103, Thr305, Tyr309 and Arg302) at positions-1, -11, -12 and -17 also conforms well with the white contours in Figure 2C.

In addition, altogether two H-bonds are observed in Figure 3 playing crucial roles in anchoring the ligand in the binding site. The -NH- of ring A as the hydrogen bond donor forms an H-bond with Thr103 with a distance of 2.28 Å. This observation correlates well with the cyan contour located around position-1 (representing the H-bond favored region) in previous CoMSIA hydrogen bond donor contour maps (Figure 2D). Besides this, another H-bond is also observed in the docking pocket, the one between the F atom at position-19 and Tyr309 (3.37 Å), which acts as a supplement for the contour map.

Our docked model also shows a comparatively large empty space around positions-11 and -12 of ring D and positions-19 and -20 of ring E, indicating that in these regions the steric interaction may be favorable. The conclusions are similar to the previous CoMSIA contour maps in Figure 2A that large blue contours show around positions-11 and -12 of ring D and positions-19 and -20 of ring E, respectively. From Figure 3A, we can see many key amino acid residues around rings C and D, which correlates well with the two yellow contours as shown in Figure 2A. Quite a lot of residues appear around position-1 creating no empty space, in contrast to the result obtained from the analysis of CoMSIA steric contour maps. The fact that position-1 plays a key role in CoMSIA field analysis (discussed in Section 2.2) may account for the difference, which in turn leads to its interaction with many residues (Trp116, Thr103, Val126 and Phe106). Several residues are also observed around rings B and C, giving supplement for the steric interactions around them, which indicates that bulky substituents around rings B and C should be harmful to the increase of binding affinity.

All in all, the docking results and the 3D contour maps complement and validate each other, indicating that the QSAR models generated in the present study are reasonable and could be used to derive useful information in the future rational design of NOP agonists.

Since our molecular docking process does not take the protein flexibility into consideration, presently 5 ns molecular dynamics simulations of NOP receptor with ligand 32 were carried out on the basis of the docked complex structure for the purpose of “drawing” a dynamic picture of the conformational changes in the NOP receptor binding site.

The RMSDs (root-mean-square deviation) of the trajectory in regard to the initial structure ranging from 2.5 to 4.5 Å are presented in Figure 4A. As a result, after 2500 ps the RMSD of the complex attains about 4.0 Å and almost remains this value for the whole process. This clearly indicates 20 metastable conformations after 2500 ps of simulation for the docked complex structure. Figure 4B depicts a superposition of the average structure for the last 1 ns and the docked structure. The right hand side picture of Figure 4B is an enlarged view of the superposition of an average structure for the last 1 ns, and the docked structure of ligand 32, where the cyan stick represents the initial structure of the docked complex and the green stick represents the MD-simulated structure, respectively. Ligand 32 is shown in blue for the initial complex and, separately, in green for the final average complex. Obviously, there are no significant changes in both the protein and ligand 32 superpositions between the docked structure and the average structure obtained from MD simulations, which verifies the reasonability of the docking model.

Discarding those compounds with unspecified agonistic activity and/or with undefined stereochemistry, a total of 103 spiropiperidines analogues with a wide spectrum of activities against the nociceptin/orphanin FQ receptor synthesized by Caldwell JP et al. were used as the dataset for molecular modeling in this study [29,30]. In vitro biological activities were converted into corresponding pK_i (−lgK_i) values and used as dependent variables in the QSAR analysis. In approximately a ratio of 4:1, all molecules was divided into training (81 compounds) and test (22) sets. The selection of the test set chemicals obeys the rule that their pK_i values are randomly but uniformly distributed in the range of the values for the whole set so that the model’s predictive power could be effectively evaluated. Table 2 shows the representative skeletons and activities of the molecules, with all structures and binding affinity values of the dataset listed in supporting information Tables S1–S3.

During the modeling process, the 3D structures of all compounds were subjected to full geometry optimization using the sketch molecule module of Sybyl 6.9 package [50]. Partial atomic charges were calculated by the Gasteiger-Huckel method [51], and energy minimizations were performed by using the Tripos force field [52] and the Powell conjugate gradient algorithm with a convergence criterion of 0.05 kcal/mol.

Molecular alignment of compounds is a crucial step in the development of 3D-QSAR models [44]. In order to obtain the best possible 3D-QSAR statistical model, two different alignment rules were adopted. The first rule is the ligand-based alignment. During the process, the most potent molecule (compound 32) was chosen as a template to fit the remaining training and test sets of compounds by using substructure-alignment function available in Sybyl. The common substructure for the alignment is described in Figure 6A, and the resulting ligand-based alignment model is shown in Figure 6B. The alignment result based on another rule, the receptor-based one, is shown in Figure 6C.

The CoMFA and CoMSIA models were generated by using Sybyl 6.9 with the default parameters. Detailed algorithms of CoMFA and CoMSIA can be easily referred to many literatures, thus we only introduce the modeling parameters in this work.

To derive the CoMFA and CoMSIA descriptor fields, a 3D cubic lattice with grid spacing of 2 Ǻ in x, y, and z directions, was generated automatically to encompass the aligned molecules. In CoMFA, the steric and electrostatic fields were calculated separately for each molecule using sp³ carbon atom probe with a charge of +1.00 and energy cut-off values of 30 kcal/mol for both the steric and electrostatic fields. The probe atom was placed at each lattice point, and its steric and electrostatic interactions with each atom in the molecule were computed using the CoMFA standard scaling.

CoMSIA similarity index descriptors were derived using the same lattice boxes as those used in CoMFA calculations. In CoMSIA, the steric, electrostatic, hydrophobic, and hydrogen-bond (H-bond) donor and acceptor descriptors were calculated using a probe atom of radius 1.0 Ǻ, charge +1.0, and hydrophobicity +1.0. A Gaussian function is used to evaluate the mutual distance between the probe atom and each molecule atom. Because of the different shape of the Gaussian function, CoMSIA similarity indices (A_F) for molecule j with atom i at grid point q are calculated by equation:

where k represents the steric, electrostatic, hydrophobic, or hydrogen-bond donor or acceptor descriptor. ω_probe,k is the probe atom with radius 1.0 Ǻ, charge +1, hydrophobicity +1, H-bond donating +1, H-bond accepting +1; ω_ik is the actual value of the physicochemical property k of atom i; γ_iq is the mutual distance between the probe atom at grid point q and atom i of the test molecule. The attenuation factor was set to 0.3.

In order to generate statistically significant 3D-QSAR models, partial least squares (PLS) regression was adopted to analyze the training set by correlating the variation in their pK_i values (the dependent variable) with variations in their CoMFA/CoMSIA interaction fields (the independent variables). PLS is a statistical approach that generalizes and combines features from principal component analysis and multiple regressions. When the matrix of predictors has more variables than observations (multicollinearity), PLS is particularly a useful way to predict a set of dependent variables from a large set of independent variables.

Leave-one-out (LOO) cross-validation analysis that one compound was moved away from the dataset and its activity was predicted by the model derived from the rest of the dataset, was performed to evaluate the reliability of the models generated from the PLS analysis. A cross-validated correlation coefficient, Q², was subsequently obtained and provided as a statistical index of the predictive power. Then, a non-cross-validation analysis was carried out with the Pearson coefficient (R²_ncv) and standard error of estimate (SEE) calculated. Finally, the CoMFA/CoMSIA results were graphically represented by field contour maps, where the coefficients were generated using the field type “Stdev*Coeff”.

In order to evaluate the real predictive ability of the best models generated by the CoMFA/CoMSIA analyses using the training set, the 26 compounds not used in the model generation are used as the external validation set. A predictive R value was then obtained with the following formula:

where SD denotes the sum of squared deviation between the biological activities of the test set molecule and the mean activity of the training set molecules; PRESS represents the sum of squared deviations between the experimental and predicted activities of the test molecules, respectively.

Molecular docking was carried out by the Surflex-dock module (V 2.51) [53] of an advanced version of Sybyl-X 1.1 [54] to understand the detailed binding model for the active site of NOP receptor with its ligands. In Surflex-docking, protomol was a computational representation of the intended binding site to which putative ligands were aligned and its construction was based on the protein residues proximal to the native ligand and on parameter settings to produce a small and buried docking target. Up to now, the protein structure has not been resoluted and the identities of our homology modeling models are below 30%. However, Luo HB from School of Pharmaceutical Sciences, Sun Yat-Sen University, who has done the similar studies on NOP receptor and used homology modeling to make a protein, is very kind to provide us his protein structure for our study. The molecular docking process is summarized as the following steps: First, the template protein structure was imported into Surflex. Then the protomol was generated using a ligand approach. Two parameters, protomol_bloat and protomol_threshold, which respectively determine how far from a potential ligand the site should extend and how deep into the protein the atomic probes used to define the protomol can penetrate, are set at 0 and 0.60 respectively. Finally, each conformer of all 103 agonists was docked into the binding site 20 times. The Hammerhead scoring function [55] is used to score the molecules in the putative poses. During the present molecular docking process, the protein was considered to be rigid, and the ligand molecules were flexible. All other parameters were setting at default values.

After docking analysis, the docked structure of compound 32 was applied in the MD simulations using the Amber 10 [56]. The general atom force field (GAFF) [57] and the standard AMBER force field for bioorganic systems (ff03) [58] were used to model the ligand and protein respectively. The docked structure was neutralized with 9 counter chloridion ions and solvated in a rectangular box of TIP3P [59] water, which kept a minimum distance of 12 Å between the solute and each face of the box (74.984 × 97.951 × 67.771 ų). The total number of the atoms of the simulation system was 40091 including the complex and waters. The cutoff distance was kept to 10 Å to compute the non-bonded interactions. All simulations were performed under periodic boundary conditions. To remove possible bad contacts, the complex was energy minimized by a multistep procedure including 9500 conjugate-gradient steps followed by 500 steepest-descent steps. Constant volume dynamics with a cutoff of 10 Å was chosen. SHAKE [60] was turned on for bonds involving H-atoms.

In the simulation process, first, the minimized system was heat up to 300 K at a constant rate of 6 K/ps while the protein atoms were constrained. The second step depended on a 50 ps pressure-constant period to raise the density and keep the complex atoms constrained. The third step was a 500 ps Langevin dynamics calculation with a collision frequency of 1 ps⁻¹, which was performed with a 2 fs time step at a constant temperature of 300 K. Finally, the production phase was run for 5 ns with a 2 fs time step. The long-range electrostatics was treated by using the particle-mesh-Ewald (PME) method [61] with default values.