QSAR models of cytochrome P450 enzyme 1A2 inhibitors using CoMFA, CoMSIA and HQSAR

Abstract

Quantitative structure–activity relationship (QSAR) studies were conducted on an in-house database of cytochrome P450 enzyme 1A2 inhibitors using the comparative molecular field analysis (CoMFA), comparative molecular similarity analysis (CoMSIA) and hologram QSAR (HQSAR) approaches. The database consisted of 36 active molecules featuring varied core structures. The model based on the naphthalene substructure alignment incorporating 19 molecules yielded the best model with a CoMFA cross validation value q² of 0.667 and a Pearson correlation coefficient r² of 0.976; a CoMSIA q² value of 0.616 and r² value of 0.985; and a HQSAR q² value of 0.652 and r² value of 0.917. A second model incorporating 34 molecules aligned using the benzene substructure yielded an acceptable CoMFA model with q² value of 0.5 and r² value of 0.991. Depending on the core structure of the molecule under consideration, new CYP1A2 inhibitors will be designed based on the results from these models.

1. Introduction

The cytochrome P450 (CYP) family of enzymes metabolizes drugs, carcinogens, steroid hormones and environmental toxicants. Of the two members of CYP1A subfamily (1A1 and 1A2), CYP1A2 is the major P450 enzyme in the human liver [1,2]. It is estimated that this enzyme metabolizes around 13% of all drugs [3–6]. CYP1A2 is primarily regulated by the aromatic hydrocarbon receptor (AhR) [7] and a basic helix–loop–helix protein belonging to the Per–Arnt–Sim family of transcription factors [8]. CYP1A2 catalyses the oxygenation of polycyclic aromatic hydrocarbons (PAHs) and heterocyclic aromatic amines/amides, the demethylation of aminoazo dyes, and the dealkylation of phenacetin, caffeine, and many other therapeutic agents [9].

Due to the role of CYP1A2 in procarcinogen activation [10,11], elevated levels of CYP1A2 could enhance individual susceptibility to carcinogenesis and modulation of this enzyme’s activity by CYP1A2 inhibitors could have important implications for cancer prevention [12,13]. A number of inhibitors for different cytochrome P450 enzymes (1A1, 1A2, 2A6, 3A4, 3A5, 1B1, 2B1, 2B4, 2B5, 2B6, 2B11, etc.) have been developed by our group, belonging to different structural classes containing adamantane, pyrene, phenanthrene, flavone, flavanone, naphthoflavone, coumarin, biphenyl, benzene, and naphthalene core structures [14–22]. Acetylenic substitution was incorporated in the majority of the inhibitors, with many of them showing mechanism-based inhibition of CYP1A1 but not of CYP1A2 [15].

The crystal structure of CYP1A2 in complex with α-naphthoflavone [23] has a rather compact active site with many phenylalanine residues (Phe125, Phe226, Phe256, Phe260 and Phe319). CYP1A2 substrates generally contain a planar ring system that can fit in the narrow and planar active site. Most of our inhibitors exhibited such a planar core structure and many of them had fused aromatic rings, which could have π–π interactions with the phenylalanine residues in the enzyme’s active site. We have previously shown that for CYP1A2, the π–π interactions with the four phenylalanine residues play a dominating role in determining the potency of the inhibitors [24]. The proximity of the triple bond to the Heme-Fe, along with favourable positioning of the hydrophobic groups of the inhibitor’s side chains positioned in the hydrophobic region of the binding pocket, contribute to increased potency of inhibition.

In addition to these insights into the structural requirements for the inhibitors, a tool such as quantitative structure–activity relationship (QSAR) would be highly beneficial in the design of new inhibitors. Different laboratories use different experimental protocols to analyse the inhibition of CYPs, which prevents direct comparison of the data obtained. The molecules and data used for this study consisted only of our in-house database since the experimental conditions used in the in vitro experiments were homogenous and the data obtained were numerically consistent, making them an optimal set for direct comparison.

QSAR, in essence, is an attempt to find a consistent relationship between the structural/chemical features of the molecules and the biological activity. We used the three-dimensional (3D) structures of the molecules, whose geometries were optimized using standard procedures available in SYBYL-X1.1. Three techniques were used for the QSAR analysis: quantitative comparative molecular field analysis (CoMFA) [25]; comparative molecular similarity analysis (CoMSIA) [26]; and hologram QSAR (HQSAR) [27].

Studies were undertaken for the 36 molecules that showed inhibition activity against CYP1A2. The basic idea of CoMFA is that the shapes of the non-covalent fields surrounding the molecules are often related to their biological property. The shape of the molecular field is incorporated in tabular form by sampling the molecule’s steric and electrostatic magnitudes at regular intervals throughout a defined region. The steric and the electrostatic properties surrounding each of the aligned molecules are calculated according to Lennard–Jones and Coulomb potentials, respectively.

CoMSIA is an extension of the CoMFA methodology and differs only in the implementation of the fields. In CoMSIA, hydrophobic, hydrogen bond acceptor and hydrogen bond donor similarity fields are calculated in addition to the steric and electrostatic fields, which allows for a better interpretation of the correlations between the 3D structures of the molecules and their activities. Partial least squares (PLS) method of analysis was our chosen method as it applies the feature extraction and induction in one step. CoMFA and CoMSIA techniques allow physical interpretation of PLS extracted model components in terms of 3D contour maps. CoMFA and CoMSIA analyses rely on the superposition of molecules so that key pharmacophore atoms (substructures) are aligned in a manner that is adequate for good results. Molecular conformation and relative alignment can be difficult and complex, especially when data sets contain structurally very diverse molecules where no obvious alignment rule suggests itself.

HQSAR is a new technique that works by utilizing substructural fragment fingerprints (molecular holograms) as predictive variables of biological activity for partial least squares analysis. The inhibition activity of each fragment allows a prediction of the inhibition effect of the molecules. By removing the requirement for molecular alignment, HQSAR models can be applied to data sets of varied sizes and is swift. The results from these studies in conjunction with our earlier docking studies will be utilized to design new inhibitors for CYP1A2 for improving their potency and selectivity.

2. Materials and methods

2.1 Compounds and software

A series of 36 compounds were used for quantitative structure-activity studies: 2-ethynylphenanthrene (2EPHEN), 3-ethynylphenanthrene (3EPHEN), 9-ethynylphenanthrene (9EPHEN), 2-(1-propynyl)phenanthrene (2MEPHEN), 9-(1-propynyl)phenanthrene (9MEPHEN), 1-ethynylpyrene (1EP), 1-propynylpyrene (1MEP), 1-butynylpyrene (1EEP), 4-propynylpyrene (4MEP), 4-methyl-7-coumarinpropargyl ether (4M7CPE), 3-(phenyl)-7-coumarinpropargyl ether (3PH7CPE), 4-(trifluoromethyl)-7-coumarinpropargyl ether (4TFM7CPE), 3,4,8-(trimethyl)-7-coumarinpropargyl ether (3,4,8TM7CPE), 2-biphenylprogargyl ether (2BIPHPE), 4-biphenylpropargyl ether (4BIPHPE), 2-biphenyl methyl propargyl ether (2BIPHMPE), 4-biphenyl methyl propargyl ether (4BIPHMPE), 4,4′-biphenyldipropargyl ether (4,4′BIPHDPE), 1-napthylpropargyl ether (1NPE), 2-napthylpropargyl ether (2NPE), 1-napthyl methyl propargyl ether (1NMPE), 2-napthyl methyl propargyl ether (2NMPE), 1-napthyl ethyl propargyl ether (1NEPE), 2-napthyl ethyl propargyl ether (2NEPE), 3′-methoxyflavone (3FM), 4′-methoxyflavone (4FM), 3′-flavonepropargyl ether (3′FPE), 4′-flavonepropargyl ether (4′FPE), 5-flavonepropargyl ether (5FPE), 6-flavonepropargyl ether (6FPE), 7-flavonepropargyl ether (7FPE), α-naphthoflavone-2′-propargyl ether (2′-αNFPE), α-naphthoflavone-4′-propargyl ether (4′-αNFPE), β-naphthoflavone-2′-propargyl ether (2′-βNFPE), β-naphthoflavone-4′-propargyl ether (4′-βNFPE) and 7-Flavonepropargyl ether (7FlavanonePE).

All studies were performed on Red Hat Linux computers running the Tripos SYBYL-X1.1 program. Initial geometric optimizations of the ligands were carried out using the standard MMFF94 force field, with a 0.001 kcal/mol energy gradient convergence criterion and a distance-dependent dielectric constant employing Gasteiger and Marsili charges. The biological activities of these compounds against CYP1A2 were measured by the assay method reported earlier [24].

2.2 Alignment of databases

The SYBYL automated alignment feature was utilized to obtain well-aligned databases. Substructural overlap wherein the molecules shared a common core of atoms was used. The naphthalene ring substructure (two fused aromatic rings) and the benzene ring substructure (monomeric aromatic ring) were employed for the alignment. The most active molecule 1MEP served as the reference for the superposition.

QSAR is based on the free energy relationship to equilibrium constants, which is proportional to the inverse of logarithm of inhibition concentration of the compound. Most frequently, for normal distribution of data, IC₅₀/EC₅₀ values are expressed as log[C] or log1/[C]. The IC₅₀ values (μM) of inhibition were expressed as 9-log(IC₅₀) for QSAR analysis.

2.3 3D QSAR through CoMFA

For the CoMFA calculations, steric and electrostatic contributions were truncated to a value of ±30 kcal/mol. All regression analyses were performed using the partial least squares algorithms in SYBYL. Initial analyses were performed using full cross-validation (leave-one-out method) and 10 principal components (PCs). In this report, the r² and q² values were measured. The r² value is the Pearson correlation coefficient, which is the correlation between the calculated activities and the observed cytotoxic activities, while q² is the predicted value based on a leave-one-out (LOO) cross-validation method [28,29]. LOO involves using one compound from the original sample as the validation data and the remaining observations as the training data. This is repeated such that each observation in the sample is used once as the validation data. After the predictive quality of the best correlation model is determined, the optimum number of components is employed to perform a non-validation PLS analysis with the same column filtering set to get the final model. The optimal number of components to be used in the noncross-validated (conventional) analyses was defined as that which yielded the highest cross-validated q² value. For component models with identical values, the component number producing the smallest standard error of prediction (SEP) was selected.

2.4 CoMSIA

The CoMFA alignment was also used to compute steric, electrostatic and hydrophobic similarity index fields for CoMSIA [26]. The steric contribution was reflected by the third power of the atomic radii of the atoms. Electrostatic properties were introduced as Gasteiger–Marsili charges. An atom-based hydrophobicity was assigned according to the parameterization developed by Viswanadhan’s group [30]. The lattice dimensions were selected with a sufficiently large margin (>4Å) to enclose all aligned molecules.

The advantage of CoMSIA fields is that no singularities occurred at atomic positions due to Gaussian-type distance dependence of the physicochemical properties. Thus, no arbitrary cutoffs were required. Similarity indices were computed using a probe with a charge of +1, a radius of +1, a hydrophobicity of +1, and 0.3 as an attenuation factor R for the Gaussian-type distance. The statistical evaluation for the CoMSIA analysis was carried out in the same way as described for CoMFA.

2.5 HQSAR

HQSAR [27] is a technique based on the concept of using molecular substructures expressed in a binary pattern (molecular hologram) as descriptors in QSAR models. The premise of HQSAR is that the two-dimensional (2D) fingerprint encodes the structure of a molecule, which is the key determinant of all molecular properties.

Two-dimensional chemical database storage and searching technologies rely on linear notations that define chemical structures (WLN – Wiswesser line-formula notation; SMILES – simplified molecular input line entry system; SLN – SYBYL line notation). Each unique fragment is assigned a specific positive integer that corresponds to a bin in an integer array of fixed length L. The process involves generation of fragments that are hashed into array bins in the range of 1 to L wherein the array is called molecular hologram and the bin occupancies are the descriptor variables. The default values for the hologram lengths are all prime numbers as this reduces the chances of seeing the same bad collisions at the various lengths. Two-dimensional similarity database searches that employ the Tanimoto coefficient between fingerprints for similarity search have used this approach with great success.

In this study, fingerprints were generated for all substructures between four and seven atoms in size for all molecules. The substructure fingerprints were then hashed into hologram bins with lengths of 53, 59, 61, 71, 83, 97, 151 and 199. LOO cross-validation was applied to determine the number of components that yields a good predictive model. PLS then yields a mathematical equation that related the molecular hologram bin values to the inhibition activity of the compounds in the database.

2.6 Progressive scrambling

Standard methods of cross-validation such as LOO can confound redundant data sets wherein the data set will have at least one ‘twin’ that is a molecule with very similar descriptor values. Cross-validation generally gives good predictions. To address the issue of false sense of confidence, the progressive scrambling method developed at Tripos [31,32] was used to determine the sensitivity of the PLS model to small systematic perturbations of the response variable. This process is done in five steps: rows are sorted with respect to the dependent variable which are then partitioned in a number of bins; the dependent variables are scrambled a number of times within each bin; each scrambling is characterized in terms of the correlation of the scrambled responses with the unperturbed data (r²yy′); SAMPLS is then applied to the perturbed data set to obtain cross-validated standard error of prediction (SDEP) and the cross-validated correlation coefficient (q²) as a function of r²yy′ at this binning level. SDEP is modelled as a function of r²yy′ at a user defined critical point (0.85) to return cSDEP (critical point). The slope of q² (the cross-validated correlation coefficient) (dq²′/dr²yy′) is evaluated at the specific critical point with respect to the correlation of the original dependent variables versus the perturbed dependent variables. The obtained CoMFA and CoMSIA analyses were subjected to the progressive scrambling process.

3. Results and discussion

A total of 36 compounds showed inhibition against CYP1A2 (Figure 1). They belonged to diverse core structural frames, including pyrenes [17,24], phenanthrenes [24], flavones [33,34], naphthoflavones [35], coumarins [36], biphenyls [37] and naphthalenes [38]. The majority of these compounds had one or more acetylene moieties in their side chains. They showed varied inhibition potencies against CYP1A2 with the pyrenes, phenanthrenes and naphthoflavones exhibiting nanomolar potencies. A good predictive model was required for this system in order to design molecules with better inhibition profiles. CoMFA, CoMSIA, and HQSAR analyses were used in the present study towards this objective.

Figure 1.

Structure of P450 enzyme 1A2 inhibitors belonging to nine core structures.

The accuracy of prediction of CoMFA models and their contour models depends to a great extent on the structural alignment [39]. Factors such as alignment rules, orientation of aligned compounds and probe atom type can influence CoMFA results. We applied molecular alignment to align all the molecules using the SYBYL routine database align option. The database was aligned using the most active compound 1MEP as an alignment template as well as other common substructures derived by taking important structural features of the inhibitors into consideration.

3.1 Alignment of databases

The active site of CYP1A2 is narrow and planar with the catalytic heme residue at one end and the group of phenylalanine residues (Phe125, Phe226, Phe256 and Phe260) at the other end. In our earlier CYP1A2 docking studies on pyrene and phenanthrene derivatives, we showed that the optimum π–π interaction of the inhibitors with the phenylalanine residues of the active site is a prerequisite for good inhibition [24]. Based on this premise, the database of 36 active compounds was aligned using naphthalene and benzene ring systems as substructures to produce aligned databases of 19 (Align1) and 34 (Align2) compounds. The ‘align database’ command in SYBYL was used to align some or all of the molecules in the database with a template molecule also in the database. A common substructure was provided to evaluate the best ‘fit’. Substructural overlap assumes that the molecules share a common core of atoms which is overlapped in each of the molecules of the database. Due to the stringent nature of this type of alignment, many compounds were excluded on using ‘naphthalene’ as substructure, which resulted in the Align1 database of 19 compounds. When the ‘benzene’ substructure was used for the alignment, most of the compounds except for the methoxyflavones were aligned giving the database Align2. Align1 consisted of compounds containing pyrene, phenanthrene, naphthoflavone, and naphthyl core structures.

A molecule has to meet certain structural and electronic configurations for it to be metabolized by the P450 enzymes. The structural characteristics will resolve its ability to fit in the active site of the enzyme and the electronic characteristics will influence the activation of oxidative metabolism. These characteristics are used by methods such as ‘COMPACT’ for predicting the potential of a chemical to bind to the cytochrome P450s [40]. One of the important structural characteristic of a compound is its logP value (P is the partition coefficient) and this property has been successfully used in QSAR [41,42]. ClogP of all the molecules in Align1 and Align2 databases was computed by the CLOGP program in SYBYL and used as an additional descriptor. These two aligned databases were subjected to CoMFA, CoMSIA and HQSAR analyses.

3.2 CoMFA, CoMSIA and HQSAR analyses of Align1 database

Align1 produced the best results for CoMFA, CoMSIA and HQSAR analyses. The summary of the output results for these analyses is given in Table 1. The CoMFA LOO analysis gave a cross-validated result of q² = 0.667 with six components. The noncrossvalidated PLS analysis resulted in a conventional r² of 0.976, F = 81.194, and a standard error of prediction (SEP) of 0.355. The steric field descriptors contributed 39.3% of the variance, the electrostatic descriptors contributed 30.1%, and ClogP descriptors contributed 30.6% (Table 3). The ClogP descriptor contributed on par with steric and electrostatic descriptors, suggesting that for the current series of molecules lipophilicity is also an important factor. The predicted inhibition, actual inhibition and their residues for database Align1 are listed in Table 2, and the correlation between the predicted inhibition and the actual inhibition are depicted in Figure 2a.

Table 1.

Summary of CoMFA, CoMSIA and HQSAR output for Align1.

	CoMFA	CoMSIA	HQSAR
Optimum no. components	6	8	6
q2	0.667	0.616	0.652
SEP	0.355	0.411	0.363
r2	0.976	0.985	0.917
s	0.095	0.082
F test	81.194	83.266
p value	0	0

Table 3.

Relative contributions of CoMFA in Align1.

Relative contributions	Fraction
Steric	0.393
Electrostatic	0.301
ClogP	0.306

Table 2.

Actual (EI) and predicted IC₅₀ (μM) (PI) values together with their residual values (δ) for CoMFA, CoMSIA and HQSAR analyses for database Align1.

	Compound	EI	CoMFA		CoMSIA		HQSAR
			PI	δ	PI	δ	PI	δ
1	1EEP	0.43	0.477	−0.047	0.461	−0.031	0.233	0.197
2	1EP	0.32	0.309	0.011	0.278	0.042	0.250	0.070
3	1MEP	0.06	0.064	−0.004	0.076	−0.016	0.238	−0.178
4	1NEPE	1.48	1.317	0.163	1.446	0.034	1.906	−0.426
5	1NMPE	3.66	3.357	0.304	3.626	0.034	1.908	1.752
6*	1NPE	1.71	1.569	0.141	1.567	0.143	1.866	−0.156
7	2EPHEN	0.34	0.548	−0.208	0.458	−0.118	0.646	−0.306
8	2MEPHEN	0.60	0.470	0.130	0.516	0.084	0.615	−0.015
9	2NEPE	1.53	1.590	−0.060	1.702	−0.172	2.043	−0.513
10	2NMPE	2.36	2.644	−0.284	2.196	0.164	2.051	0.309
11	2NPE	3.51	3.490	0.020	3.508	0.002	2.067	1.443
12	2′αNFPE	1.38	1.426	−0.046	1.485	−0.105	0.324	1.056
13*	2′βNFPE	0.51	0.514	−0.004	0.580	−0.070	0.218	0.292
14	3EPHEN	0.28	0.226	0.054	0.218	0.062	0.640	−0.360
15	4MEP	0.45	0.456	−0.006	0.446	0.004	0.218	0.232
16	4′αNFPE	0.26	0.244	0.017	0.243	0.017	0.368	−0.108
17	4′βNFPE	0.07	0.073	−0.003	0.064	0.006	0.253	−0.183
18*	9EPHEN	0.40	0.532	−0.132	0.460	−0.060	0.553	−0.153
19	9MEPHEN	0.67	0.490	0.180	0.579	0.091	0.521	0.149

^*Compounds that were not included in the training set of the 3D-QSAR model.

Figure 2.

Relationship between predicted IC₅₀ (μM) from CoMFA (a), CoMSIA (b) and HQSAR (c) models and the experimental IC₅₀ (μM) values of P450 enzyme 1A2 inhibitors. Data points representing testing compounds that were not included in the training set of QSAR analyses are coloured red.

The summary of the CoMSIA analysis of database Align1 is shown in Table 1. CoMSIA results gave a cross-validated q² of 0.616 for eight components and non-crossvalidated r² of 0.9785, F = 83.266 and SEE = 0.411. The corresponding field contributions of steric, electrostatic, hydrophobic, hydrogen-bond acceptor, hydrogen-bond donor+acceptor (steric), steric+electrostatic (steric) and steric+electrostatic (electrostatic) were 7.0%, 29.3%, 10.8%, 8.2%, 8.2%, 7.0% and 29.3%%, respectively (Table 4). The result indicates that the electrostatic field is very important for the present series of molecules. The predicted inhibition, actual inhibition and their residues for database Align1 are listed in Table 2. Figure 2b shows the relationship between the predicted and the experimental IC₅₀ values for the CoMSIA model.

Table 4.

Relative contributions of CoMSIA in Align1.

Relative contributions	Fraction
Steric	0.070
Donor and acceptor (steric)	0.082
Steric and electrostatic (steric)	0.070
Steric and electrostatic (electrostatic)	0.293
Acceptor	0.082
Hydrophobic	0.108
Electrostatic	0.293

Table 1 summarizes the results of the HQSAR calculation for Align1. The lowest standard error occurred at a cross-validated q² of 0.652 with six optimal components. The hologram that gave the lowest standard error had a hologram length of 83. The PLS analysis yielded a conventional r² of 0.917 and a standard error prediction of 0.363 for all of the studied compounds. The predicted inhibition, actual inhibition and their residues for database Align1 are listed in Table 2, and the correlation of the HQSAR predicted values vs. the actual values is presented graphically in Figure 2c. It is important to have a QSAR technique that offers not only a consistent and reproducible prediction but also a fast and convenient procedure. These results indicate that HQSAR is fairly capable of predicting the effective inhibition accurately and quickly. The 2D fingerprint generated by the HQSAR analysis will be used in future database searches for new CYP1A2 inhibitors.

Table 1 clearly shows that all statistical data for the CoMFA, CoMSIA and HQSAR analyses for Align1 have high predictability and stability. The Pearson correlation coefficient, r², shows how much the predicted activity approximated the inhibition activity. CoMFA and CoMSIA showed comparable r² of 0.976 and 0.985 and hence both can be utilized as predictive models. HQSAR also depicted a reasonable r² value of 0.917, indicating that it is less predictive than CoMFA and CoMSIA. Three compounds (marked by an asterisk in Table 2) were selected to test the predicting capability and were not included in the process of the construction of the above QSAR models. The results are shown in Table 2 and Figure 2, and the predicted IC₅₀ values were consistent with the experimental data in the tolerable error range.

3.3 CoMFA, CoMSIA and HQSAR analyses of Align2 database

The Align2 database, which included all the active compounds in its alignment, was also subjected to CoMFA, CoMSIA and HQSAR analyses (Table 5). CoMFA analysis gave a cross-validated result of q² = 0.5 for eight components which was in the acceptable range. The Pearson correlation coefficient r² showed good predictability of 0.991 with F = 334.36 and SEP = 0.369. Relative contributions (Table 6) were evidenced by the descriptors ClogP (19.1%), hydrogen bond (steric) (20%), indicator (21.8%), steric (17.3%) and electrostatic (21.8%). Analogous to CoMFA analysis on Align1, we found that ClogP descriptor made an equivalent contribution to the model implicating that lipophilicity is a key factor for CYP1A2 inhibitors. All these descriptors seem to contribute equally to the predictive model. The actual inhibition, predicted inhibition and their deviations are listed in Table 7, and graphically represented in Figure 3. CoMSIA and HQSAR analyses results (Table 5) indicated that these analyses did not yield good predictive models. The data clearly showed that the CoMFA result was significantly superior to CoMSIA and HQSAR analyses results for Align2.

Table 5.

Summary of CoMFA, CoMSIA and HQSAR output for Align2.

	CoMFA	CoMSIA	HQSAR
Optimum no. components	8	3	1
q2	0.500	0.175	0.439
SEP	0.369	0.433	0.343
r2	0.991	0.966	0.553
s	0.050	0.093
F test	334.360	126.374
p value	0	0

Table 6.

Relative contributions of CoMFA to Align2.

Relative contributions	Fraction
ClogP	0.191
HBond (steric)	0.200
IND (steric)	0.218
FXCOMFA (steric)	0.173
FXCOMFA (electrostatic)	0.218

Table 7.

Actual (EI) and predicted IC₅₀ (μM) (PI) values together with their residual values (δ) of CoMFA analyses for database Align2.

	Compound	EI	CoMFA
			PI	δ
1	1EEP	0.43	0.45	−0.02
2	1EP	0.32	0.25	0.07
3	1MEP	0.06	0.06	0.00
4	1NEPE*	1.48	1.43	0.05
5	1NMPE	3.66	3.59	0.07
6	1NPE	1.71	1.70	0.01
7	2BIPHMPE	1.73	1.55	0.18
8	2BIPHPE	5.80	6.08	−0.28
9	2EPHEN	0.34	0.47	−0.13
10	2MEPHEN	0.60	0.54	0.06
11	2NEPE	1.53	1.67	−0.14
12	2NMPE	2.36	2.27	0.09
13	2NPE	3.51	3.32	0.19
14	2′αNFPE	1.38	1.36	0.02
15	2′βNFPE*	0.51	0.54	−0.03
16	348TM7CPE	0.44	0.45	−0.01
17	3EPHEN*	0.28	0.28	0.00
18	3PH7CPE	1.65	1.67	−0.02
19	3′FPE	0.67	0.66	0.01
20	44′BIPHDPE	3.19	3.32	−0.13
21	4BIPHMPE*	1.84	1.77	0.07
22	4BIPHPE	1.55	1.42	0.13
23	4M7CPE	2.01	2.08	−0.07
24	4MEP	0.45	0.49	−0.04
25	4TFM7CPE	0.70	0.72	−0.02
26	4′αNFPE	0.26	0.26	0.00
27	4′βNFPE	0.07	0.07	0.00
28	4′FPE	1.27	1.30	−0.03
29	5FPE	0.38	0.34	0.04
30	6FPE*	0.94	0.93	0.02
31	7FLAVANONEPE	0.97	0.97	0.00
32	7FPE	0.39	0.41	−0.02
33	9EPHEN	0.40	0.51	−0.11
34	9MEPHEN	0.67	0.57	0.10

^*Compounds that were not included in the training set of the 3D-QSAR model.

Figure 3.

Relationship between predicted and actual IC₅₀ (μM) values from CoMFA model for database Align2. Data points representing testing compounds that were not included in the training set of QSAR analyses are triangles coloured red.

3.4 Progressive scrambling

To determine the stability of these models, we performed progressive scrambling on the CoMFA and CoMSIA analyses of the models based on Align1 and Align2. The progressive scrambling data is shown in Table 8. Progressive scrambling of the CoMFA model of Align1 gave the best correlation at six components, with q² = 0.486 and dq²′/dr²yy′ = 1.044. For the CoMSIA model of Align1, progressive scrambling performed the best analysis with eight components, with q² = 0.420 and dq²′/dr²yy′ = 0.745. In the case of CoMFA model for Align2, the highest q² = 0.290 was found for eight components, but the dq²′/dr²yy′ = 1.200 showed that it had a less acceptable slope. The entry with the best slope dq²′/dr²yy′ = 1.100 was for seven components with a q² = 0.273. This showed that the Align2 model was not a perfect one, but it can be used with caution. The progressive scrambling of the CoMFA and CoMSIA PLS of Align1 and Align2 demonstrated that these models are not dependent on chance correlation.

Table 8.

Results of progressive scrambling of CoMFA and CoMSIA analyses.

Components	Align1 CoMFA			Align1CoMSIA			Align2 CoMFA
	q2	cSDEP	dq2/dr2yy′	q2	cSDEP	dq2/dr2yy′	q2	cSDEP	dq2/dr2yy′
3	0.225	0.479	0.845	0.150	0.504	0.522	0.151	0.439	0.579
4	0.357	0.453	0.738	0.132	0.524	0.264	0.181	0.438	0.708
5	0.450	0.432	0.835	0.289	0.494	0.470	0.271	0.420	0.750
6	0.486	0.436	1.044	0.312	0.505	0.659	0.262	0.430	0.951
7	0.464	0.466	0.900	0.389	0.500	0.677	0.273	0.435	1.100
8	0.458	0.492	0.800	0.420	0.510	0.745	0.290	0.438	1.200
9	0.453	0.521	0.769	0.391	0.555	0.652	0.280	0.450	1.310
10	0.452	0.553	0.760	0.360	0.603	0.399	0.274	0.462	1.331
15	0.455	0.900	0.761	0.332	1.004	0.324	0.386	0.517	1.312

3.5 Contour maps

The CoMFA 3D coefficient contour maps for the Align1 database are shown in Figure 4a and 4b for the most active compound 1-propynylpyrene (1MEP) and the least active compound 1-napthyl methyl propargyl ether (1NMPE), respectively. Beneficial and detrimental steric interactions are displayed in green and yellow contours, respectively, while red and blue contours illustrate regions of desirable negative and positive electrostatic interactions. The fused aromatic rings of the inhibitors clearly lay in the favourable electrostatic region and the methyl end group of 1MEP (Figure 4a) lies in the sterically favoured region – perhaps contributing to its high inhibition potency. In contrast, the side chain of 1NMPE (Figure 4b) terminates in the sterically unfavourable region suggesting a reason for its reduced potency.

Figure 4.

(a) and (b) CoMFA contour maps for the most active (1MEP) and least active (1NMPE) compounds of database Align1, respectively. (c), (d) and (e) CoMSIA contour maps for the most active (1MEP), least active (1NMPE) and intermediate active (4′βNFPE) compounds of database Align1, respectively.

The CoMSIA results for Align1 indicate that electrostatic descriptors are the major contributors for this model. A look at the CoMSIA contour maps for electrostatic and steric descriptors (with same colour coding as for CoMFA contour maps) in Figures 4c and 4d for the compounds 1MEP and 1NMPE clearly show the presence of favourable electrostatic effects for the fused aromatic rings. Similar to the CoMFA contour map, the side chain end methyl group of 1MEP (Figure 4a) resides in the sterically favoured region contributing to its high potency and the side chain of 1NMPE (Figure 4b) culminates in a sterically unfavourable region explaining its lowered potency.

Closer inspection of the steric and electrostatic regions of the CoMFA and CoMSIA contour maps of 1MEP shows that most of these regions overlap. Figure 5 illustrates the CoMFA and CoMSIA electrostatic and steric maps superimposed on the structure of the most active compound in the training set. Comparison of the contour maps from CoMFA and CoMSIA showed high level of compatibility with the active site of CYP1A2 thereby reinforcing our confidence in this model. The favoured electrostatic regions all occur over the aromatic rings of the compounds (as seen for 4′βNFPE in Figure 4e) reasserting our earlier conclusion that π–π interactions between the enzyme aromatic residues and the inhibitor aromatic residues are one of the main determinants of the inhibitor’s potency. The distances between the centroids of the nearest aromatic rings of compound 1MEP and the aromatic ring of phenylalanine residues of the enzyme (Figure 5) are 1MEP-PHE125 = 5.7 Å, 1MEP-PHE226 = 3.2 Å, 1MEP-PHE256 = 5.7 Å, and 1MEPPHE260 = 5.3 Å.

Figure 5.

Binding mode of 1MEP to CYP1A2 active site. The superimposition of the (a) CoMFA and (b) CoMSIA electrostatic and steric plots with the structural model of the binding site is shown. The sterically favoured regions are shown in green and the sterically disfavoured regions are shown in yellow. The positive electrostatic contours are shown in blue and the negative electrostatic contours are shown in red. The phenylalanine residues having π–π interactions with the ligand 1MEP are shown in orange.

The contour maps from CoMFA and CoMSIA analyses can be used in tandem for designing new inhibitors for core structures in Align1 database. For example, in the case of the pyrenes, bulky substitutions at position 5, 9 and 10 should be avoided as they reside in unfavourable steric regions (CoMFA map Figure 4a). In the CoMSIA (Figure 4c) contour map of 1MEP, the blue contour indicating an electrostatically favourable region appears near position 4, 5 and 6. If we avoid position 5 due to steric considerations, positions 4 and 6 seem suited for electronegative substituents that should increase its inhibition potency.

4. Conclusions

Two CoMFA models were developed for the inhibitors of cytochrome P450 enzyme 1A2. The first model had a high predictability for all three analyses (CoMFA, CoMSIA and HQSAR) giving good values for LOO cross validation (q²) and for Pearson correlation coefficient (r²). Compounds containing two fused aromatic rings (naphthalene substructure) were the only ones that aligned in this database. This model can be reliably used for compounds with the pyrene, phenanthrene, naphthoflavone and naphthalene core structures. In fact, we can clearly see that the compounds with the best inhibition potencies are present in this database. For the flavones, coumarins and biphenyls the second model would be the best suited since this CoMFA model gives a reasonable predictability.

QSAR studies in conjunction with our earlier docking studies have clearly shown the greater dependence of inhibition potency on the presence of two or more fused aromatic ring substructural features in the inhibitors. These models can serve as a guide for designing new compounds belonging to the above given classes of core structures with appropriate functional group substitutions based on the CoMFA and CoMSIA derived contour maps. The 2D fingerprints derived from HQSAR can be used to search chemical databases for new molecules as potential inhibitors of CYP1A2. These studies will help us design new inhibitors to achieve increased potency and selectivity for CYP1A2.