Pharmacophore modelling and atom-based 3D-QSAR studies on N-methyl pyrimidones as HIV-1 integrase inhibitors

Abstract

Pharmacophore modelling and atom-based 3D-QSAR studies were carried out for a series of compounds belonging to N-methyl pyrimidones as HIV-1 integrase inhibitors. Based on the ligand-based pharmacophore model, we got 5-point pharmacophore model AADDR, with two hydrogen bond acceptors (A), two hydrogen bond donors (D) and one aromatic ring (R). The generated pharmacophore-based alignment was used to derive a predictive atom-based 3D-QSAR model for the training set (r² = 0.92, SD = 0.16, F = 84.8, N = 40) and for test set (Q² = 0.71, RMSE = 0.06, Pearson R = 0.90, N = 10). From these results, AADDR pharmacophore feature was selected as best common pharmacophore hypothesis, and atom-based 3D-QSAR results also support the outcome by means of favourable and unfavourable regions of hydrophobic and electron-withdrawing groups for the most potent compound 30. These results can be useful for further design of new and potent HIV-1 IN inhibitors.

Keywords::

• 3D-QSAR
• AIDS
• HIV-1 IN
• N-methyl pyrimidones
• pharmacophore modelling

Introduction

Acquired immunodeficiency syndrome (AIDS) is a disease of the human immune system caused by the human immunodeficiency virus (HIV), which infects primarily vital cells such as helper T cells, macrophages and dendritic cells^1,2. AIDS was first reported in 1981 and subsequently isolated in 1983^3–5. The virus encodes three enzymes: reverse transcriptase (RT), protease (PR) and integrase (IN). Combination antiviral therapy with RT and PR inhibitors has been exploited for better treatment of AIDS. However, the ability of HIV to rapidly develop resistance along with toxicity problem demands further discovery of novel classes of anti-HIV drugs^6–8. Recently, HIV-1 IN has emerged as a promising target for antiretroviral therapy as this is responsible for the integration of the newly synthesized double-stranded viral DNA into the host genomic DNA^8–10.

The integration of HIV-1 DNA into the host chromosome comprises a series of DNA-cutting and -joining reactions. The first step in the integration process is 3′-end processing. In this step, two nucleotides are removed from each 3′-end of the viral DNA and expose 3′-hydroxyl group. In further step of DNA strand transfer, the previously processed viral DNA end is inserted into the target DNA^11–13. Thus, IN is essential for viral replication and represents a potential target for antiretroviral drug design¹⁴. HIV-1 IN is composed of three structurally and functionally distinct domains: an N-terminal domain (amino acids 1–49) effectively binds zn²⁺, the catalytic core domain (amino acids 50–212) encompassing a DDE motif which is responsible for the catalytic activity and the C-terminal domain (amino acids 213–288) binds non-specifically to DNA^15–17.

The aim of this study is to develop pharmacophore of N-methyl pyrimidones series of HIV-1 IN inhibitors and also to build atom-based 3D-QSAR model for better understanding of structural requirement, which can be used for lead optimization and virtual screening. We used PHASE module of Schrödinger suite to develop pharmacophore models and 3D-QSAR models of HIV-1 IN inhibitors. PHASE performs systematic explorations about rotatable bonds and calculates the associated conformational energies, retaining only the most reasonable conformations. It quickly locates pharmacophores using a high-dimensional, tree-based partitioning algorithm in which pharmacophores from different conformations are placed in multidimensional boxes. After PHASE has located reasonable alignment of active ligands, generated pharmacophores are evaluated according to an open, highly configurable scoring function. PHASE determines how molecular structure effects drug activity by dividing space into a fine cubic grid, encoding atom type occupation as numerical information and performing a partial least squares (PLS) regression, resulting in the prediction of a significant model^18–20.

Materials and methods

PHASE 3.2 implemented in the Maestro 9.1 software package^18–21 was used to generate pharmacophores and 3D-QSAR models for HIV-1 IN inhibitors.

Data set

For this study, 49 compounds of N-methyl pyrimidones²² and raltegravir²³ having experimental activities are used to predict theoretical activity are shown in Tables 1 and 2. The biological activity data was reported in the form of IC₅₀. The IC₅₀ values were converted into pIC₅₀ using the formula (pIC₅₀= −log IC₅₀). We divided the data set randomly choosing 40 compounds in training set and 10 compounds in test set to maintain the 4:1 ratio. While dividing the training set and test set, we ensured the uniform distribution of structurally different compounds with a wide range of pIC₅₀ value in both training and test set.

Table 1.  Structures and actual versus predicted pIC₅₀ of compounds 1-39.

Compound	R	Actual pIC50	Predicted pIC50	Residuals
1*		7.20	7.32	–0.12
2		6.14	6.22	–0.08
3		8	7.83	0.17
4		6.16	6.18	–0.02
5		7	7.19	–0.19
6		6.74	7.04	–0.3
7		6.69	6.86	–0.17
8		6.88	6.79	0.09
9		7.52	7.61	–0.09
10		7.21	7.34	–0.13
11		7.69	7.36	0.33
12		7.69	7.35	0.34
13		7.52	7.57	–0.05
14		6.67	6.76	–0.09
15		7.67	7.57	0.1
16		7.20	7.46	–0.26
17		7.22	7.34	–0.12
18		7.15	7.19	–0.04
19*		7.43	7.25	0.18
20*		6.85	6.87	–0.02
21		7	6.86	0.14
22		7.30	7.13	0.17
23		6.69	6.77	–0.08
24*		6.60	6.87	–0.27
25		6.88	6.90	–0.02
26		7.38	7.35	0.03
27		7.49	7.39	0.1
28		5.39	5.39	0
29		7.92	7.90	0.02
30		8.39	8.49	–0.1
31		7.28	7.21	0.07
32		7.95	7.95	0
33		8.15	8.20	–0.05
34*		8.09	7.86	0.23
35		7.30	7.37	–0.07
36		7.48	7.48	0
37		7.82	8.00	–0.18
38*		7.69	7.54	0.15
39*		7.82	7.30	0.52

*Compounds used for test set.

Table 2.  Structures and actual versus predicted pIC₅₀ of compounds 40–50.

Compound	R1	R2	N	Actual pIC50	Predicted pIC50	Residuals
40	H	H	1	7.18	7.27	–0.09
41	H	H	2	7.14	7.01	0.13
42	2-F	H	1	6.75	6.87	–0.12
43*	3-F	H	1	7.20	7.29	–0.09
44	3-Cl	H	1	7.20	7.44	–0.24
45	3-Br	H	1	7.72	7.31	0.41
46*	3-OMe	H	1	6.85	7.04	–0.19
47*	4-F	3-F	1	7.67	7.36	0.31
48	4-F	3-Me	1	7.06	7.04	0.02
49	4-F	3-Cl	1	7.55	7.51	0.04
50	4-F	3-Br	1	7.35	7.23	0.12

*Compounds used for test set.

Ligand preparation

Ligprep module incorporated in PHASE was used to convert 2D to 3D structures, and minimization was performed using OPLS 2005 force field²⁴, with implicit GB/SA solvent model. The conformers were generated using rapid torsion angle approach, and structures with high estimated energies are eliminated. A maximum of 1,000 conformers were generated for each structure using preprocess minimization of 100 steps and postprocess minimization of 50 steps. Each minimized conformer was filtered through a relative energy window of 10 kcal mol⁻¹ and a minimum atom deviation of 1.00 Å^18–21. This value (10 kcal mol⁻¹) sets an energy threshold relative to the lowest energy conformer. Conformers having higher energy than threshold are discarded. All distances between pairs of corresponding heavy atoms must be below 1.00 Å for two conformers to be considered identical. This criterion is applied only after the energy difference threshold and only if two conformers are within 1 kcal mol⁻¹.

Generation of pharmacophoric sites

The six built-in pharmacophore features, hydrogen bond acceptor (A), hydrogen bond donor (D), hydrophobic group (H), negatively charged group (N), positively charged group (P) and aromatic ring (R), were used to create pharmacophore sites. The pharmacophore feature is defined by a set of chemical structure patterns. The defined patterns are specified as SMARTS queries¹⁸ and assigned one of three possible geometries, which define physical characteristics of the site: (1) Point—the site is located on a single atom in the SMARTS query; (2) Vector—the site is located on a single atom in the SMARTS query and will be assigned directionality according to one or more vectors originating from the atom; and (3) Group—the site is located at the centroid of a group of atoms in the SMARTS query. For aromatic rings, the site is assigned directionality defined by a vector that is normal to the plane of the ring. A default setting having acceptor (A), donor (D), hydrophobic (H), negative (N), positive (P) and aromatic ring (R) was used to create pharmacophore sites. The active and inactive ligands are defined by clicking activity thresholds. The active and inactive that thresholds are set 7.90 and 6.20, which yield six actives and three inactives were used for pharmacophore modelling and subsequent scoring.

Finding common pharmacophore and scoring hypotheses

Common pharmacophoric features were identified from a set of variants—a set of features that define a possible pharmacophore using a tree-based partitioning algorithm. The terminal size of the box was 1Å, which governs the tolerance on matching—the more closely related pharmacophores having smaller box sizes. Common pharmacophore hypotheses were generated by varying the number of sites (n_sites) and the number of matching active compounds (n_act). We used n_act = n_{act_tot} initially (n_{act_tot} is the total number of active compounds in the training set), n_sites was varied from seven to three until at least one hypothesis was found and scored successfully. The common pharmacophore hypotheses were scored by setting the root mean square deviation (RMSD) value below 1.0, the vector score value to 0.5 and weighing to include consideration of the alignment of inactive compounds using default parameters.

Building the 3D-QSAR model

Presently, the pharmacophore modelling and 3D-QSAR studies are successfully employed in drug design^25–28. The main purpose of applying 3D-QSAR study is to reveal the necessary features and properties required for structure of ligands and their biological activities. All the common pharmacophore hypotheses generated were used to generate atom-based 3D-QSAR models by correlating the actual and predicted activity for the set of 40 training molecules using PLS analysis. The PLS regression was carried out using PHASE with a maximum of N/5 PLS factors (N = number of ligands in training set, and a grid spacing of 1.0 Å). All models were validated by predicting activity of the set of 10 test molecules.

Results and discussion

Determination of pharmacophore and 3D-QSAR models

A pharmacophore modelling and 3D-QSAR studies were performed successfully on N-methyl pyrimidones to understand the effect of spatial arrangement of structural features on HIV-1 IN inhibition. Six compounds with highest activity were selected for common pharmacophore hypotheses generation. Using a tree-based partition algorithm requiring all six active compounds should match, 5 probable common pharmacophore hypotheses were generated from the list of variants, based on five pharmacophoric features. On applying the scoring function for five featured pharmacophore hypotheses using default values, three best common pharmacophore hypotheses, namely, AADDR, AADRR and AADHR, were selected for 3D-QSAR model building. Training set compounds were aligned on these common pharmacophore hypotheses and are analysed in PHASE with five PLS factors. The predictivity of each hypothesis was re-evaluated by the test set compounds. A summary of the statistical data for the three common pharmacophore hypotheses of AADDR, AADRR and AADHR is listed in Table 3.

Table 3.  Quantitative Structure Activity Relationship (QSAR) results for the three best Common Pharmacophore Hypotheses (CPHs).

	AADDR	AADRR	AADHR
SD	0.16	0.13	0.13
r^2	0.92	0.94	0.95
F	84.8	127.5	129.5
P	3.294e–018	5.129e–021	4e–021
RMSE	0.06	0.38	0.33
Q^2	0.71	0.29	0.46
Pearson R	0.90	0.61	0.69

SD =  Standard deviation of the regression, r² = correlation coefficient, F =  variance ratio, p = significance level of variance ratio, RMSE = root mean-square error, Q² = predictive coefficient of the test set, Pearson R = Correlation between the predicted and observed activity for the test set.

The statistical parameters r², Q², SD, RMSE and F were used to evaluate the good-quality QSAR model. The three models show good and consistent r² greater than 0.91, SD values lesser than 0.2 are high F-test values. This shows that these three models interpreting structure activity relationship for this series of training set compounds satisfactorily. According to Tropsha²⁹, high r² is a necessary but not sufficient condition for a QSAR model. Besides the consideration of high r², the best QSAR model should be chosen based on its predictive ability. The AADDR shows good external predictive ability Q² value 0.71. The AADDR hypothesis has highest Q² value in comparison with other two pharmacophore hypotheses, suggesting AADDR hypothesis is the best model among three. Additionally, AADDR pharmacophore hypothesis has low RMSE value and highest Pearson R value of all the pharmacophore hypotheses, which also supports this hypothesis. Finally, based on r², Q², SD, RMSE and Pearson R, the best model was AADDR hypothesis.

Common pharmacophore model alignment of active compounds is shown in Figure 1. The plots of actual versus predicted pIC₅₀ are shown in Figure 2 for training set and test set compounds. Distances and angles of pharmacophore hypothesis AADDR are given in Supplementary Table 1.

Figure 1.  Common pharmacophore model alignment of active compounds. Spheres with vectors A4 and A5 are acceptor features, sphere R11 is aromatic ring feature and spheres with vectors D7 and D8 are represents donor features.

Figure 2.  Scatter plot for the predicted and actual pIC₅₀ values for AADDR hypothesis applied to (a) the training set (r² = 0.92, SD = 0.16, F = 84.8, N = 40) and (b) the test set (Q² = 0.71, RMSE = 0.06, Pearson R = 0.90, N = 10).

QSAR visualization

One of the major advantages of PHASE 3D-QSAR technique is to get contour cubes based on favourable and unfavourable regions, which could be visualized in 3D space. The contour cubes obtained from AADDR show how 3D-QSAR methods can identify features which is important for the interaction between ligands and their target protein. Such contour cubes allow identification of those positions that require a particular physicochemical property to enhance the bioactivity of a ligand. A pictorial representation of the contours generated is shown in supplementary Figure 1. In these representations, blue cubes indicate favourable regions whereas red cubes indicate unfavourable regions to enhance the activity. These cubes can be generated for different properties such as hydrophobic, hydrogen bond donor, hydrogen bond acceptor (electron withdrawing) and positive and negative ionic features, which define the non-covalent interactions with receptor.

Visualization of QSAR model cubes associated with hydrophobic property gives an idea about topology of receptor site. Supplementary Figures 1A and 1B show cubes generated for hydrophobic property using QSAR model. It illustrates the significant favourable regions and unfavourable hydrophobic/non-polar interactions that arise when the QSAR model is applied to the most active compound 30 and most inactive compound 28. Blue cubes were seen on the most active compound 30 at the N, N-dimethyl position. This clearly indicates that hydrophobic substitutions can be accepted at this position to increase the activity.

Supplementary Figures 1C and 1D compare the most significant favourable and unfavourable hydrogen bond acceptors or electron withdrawing features that arise, when the QSAR model is applied to the most active compound 30 and to the most inactive compound 28. Electron withdrawing groups (N and O) associated with blue-coloured cubes in the case of most active compound 30 are observed near atoms shows high electron-withdrawing feature, whereas in the case of most inactive compound 28, red-coloured cubes were visible indicating activity difference in these two molecules. In Supplementary Figure 1, blue cubes represents the sterically favoured spatial regions to enhance the activity, whereas red cubes represent the sterically unfavoured regions.

Conclusion

Several pharmacophore hypotheses of HIV-1 IN inhibitors were developed using PHASE, and alignment based on these pharmacophores was used as input for the development of atom-based 3D-QSAR model. A five-point pharmacophore with two hydrogen bond acceptors (A), two hydrogen bond donor (D) and one aromatic ring (R) as pharmacophoric features were associated with a 3D-QSAR model giving good statistical significance and predictive ability. Visualization of the 3D-QSAR model provides details of relationship between structure and activity among these molecules and thus provides explicit indications for the design of better analogues. Furthermore, the activity-based cubes generated, using 3D-QSAR model along with finally obtained pharmacophoric features, strongly communicates picture of probable active site of the target and can be used as a useful tool for the rational drug-design process. The results of this study are expected to be useful for balanced modification of ligands as potential HIV-1 IN inhibitors with good anti-HIV activity, which can be achieved by incorporating predicted structural features to the molecule. Finally, atom-based 3D-QSAR model presented here could be very useful for the development of new leads as HIV-1 IN inhibitors.

Acknowledgements

The authors thank M. Ravikumar for the detailed discussion and suggestions. The authors thank an anonymous referee for the valuable suggestion.

Declaration of interest

The authors report no declarations of interest.