MOLECULAR CHARACTERIZATION OF CYP2B6 SUBSTRATES

Abstract

CYP2B6 has not been as fully characterized at the molecular level as other members of the human cytochrome P450 family. As more widely used in vitro probes for characterizing the involvement of this enzyme in the metabolism of xenobiotics have become available, the number of molecules identified as CYP2B6 substrates has increased. In this study we have analyzed the available kinetic data generated by multiple laboratories with human recombinant expressed CYP2B6 and along with calculated molecular properties derived from the ChemSpider database, we have determined the molecular features that appear to be important for CYP2B6 substrates. In addition we have applied 2D and 3D QSAR methods to generate predictive pharmacophore and 2D models. For 28 molecules with K_m data, the molecular weight (mean ± SD) is 253.78±74.03, ACD/logP is 2.68±1.51, LogD_{pH 5.5} is 1.51±1.43, LogD_{pH 7.4} is 2.02±1.25, hydrogen bond donor (HBD) count is 0.57 ±0.57, hydrogen bond acceptor (HBA) count is 2.57±1.37, rotatable bonds is 3.50±2.71 and total polar surface area (TPSA) is 27.63±19.42. A second set of 15 molecules without K_m data possessed similar mean molecular property values. These properties are comparable to those of a set of 21 molecules used in a previous pharmacophore modeling study (Ekins et al., J Pharmacol Exp Ther 288 (1), 21–29, 1999). Only the LogD and HBD values were statistically significantly different between these different datasets. We have shown that CYP2B6 substrates are generally small hydrophobic molecules that are frequently central nervous system active, which may be important for drug discovery research.

INTRODUCTION

The cytochrome P450 (CYP) enzymes are ubiquitous with important roles in human xenobiotic metabolism. One of these, CYP2B6 has been virtually ignored by pharmaceutical researchers for many years. CYP2B6 is expressed in hepatic and extra hepatic tissues and it has recently been suggested as a prognostic factor for prostate cancer and may therefore be clinically relevant [1]. Until recently, few selective substrates, inhibitors and monoclonal antibodies were available for analysis of CYP2B6 function [2–4]. However, the improved availability of tools (such as bupropion as a model substrate [5]) enabled more studies that characterized and quantified the involvement of CYP2B6. CYP2B6 has since been shown to have a key role in the metabolism of several important marketed compounds including efavirenz [6] and others [4] (Table 1–3). To date the three-dimensional structure of human CYP2B6 is unknown unlike CYP2C9, CYP2D6 and CYP3A4 [7, 8] although the structure of the rabbit CYP2B4 is also available [9, 10]. Efforts have been made to use available CYP2B6 substrates to predict the key molecular features necessary for interaction with the enzyme. In the years since the first validated pharmacophore and 3D-quantitative structure activity relationship (3D-QSAR) analyses for CYP2B6 indicated the importance of hydrophobic features and at least one hydrogen bond acceptor (HBA) [5], additional QSAR methods and technological developments have been applied to substrates [11, 12] and inhibitors [13]. Homology models have also been developed [11, 14–16] that used CYP2C5 (50% sequence identity with CYP2B6 [16]) and more recently CYP2B4 [12] which has 77% sequence identity with CYP2B6 (Fig 1), generally confirming the substrate pharmacophore models. We are also now acutely aware that there are several genetic polymorphisms in CYP2B6 affecting gene expression and/or enzyme activity which are clinically relevant [17]. Induction of CYP2B6 expression may also occur by activation of the constitutive androstane and pregnane×receptors [18]. Both genetic variation and enzyme induction contribute to the wide variability of CYP2B6 protein levels (~100 fold) and activity previously noted [3, 19].

Table 1.

Molecules with K_{m (apparent)} data derived from recombinant expressed CYP2B6 and molecular property data from ChemSpider.

Substrate	Km (µM)	Metabolic Reaction	Reference	MW	ACD/logP	LogDpH5.5	LogDpH7.4	HBD	HBA	RBN	PSA
Nevirapine	834	Hydroxylation	[54]	266.30	1.50	1.48	1.5	1	5	1	49.33
(R)-methadone	41	N-demethylation	[37]	309.44	4.20	1.23	2.56	0	2	7	20.31
(S)-methadone	21	N-demethylation	[37]	309.44	4.20	1.23	2.56	0	2	7	20.31
N-methyl,N-propargyl-2-phenylethylamine	290	N-demethylation	[55]	173.25	2.61	0.68	2.28	0	1	4	3.24
(+)-fenchone	180	Hydroxylation	[56]	152.23	2.13	2.13	2.13	0	1	0	17.07
Endosulfan Alpha	16.2	Sulfation	[57]	406.92	3.12	3.12	3.12	0	3	0	54.74
Prasugrel	2.3	Ring opening	[58]	373.44	2.87	2.85	2.87	0	4	6	74.85
NNK	33	Hydroxylation	[59]	207.23	0.09	0.09	0.09	0	5	6	62.63
(S)-Nicotine	820	Hydroxylation	[59]	162.23	0.72	−2.07	−0.47	0	2	1	16.13
Sertraline	30.7	N-demethylation	[60]	306.23	4.81	1.77	2.77	1	1	2	3.24
Meperidine	356	N-demethylation	[61]	247.33	2.35	−0.44	1.15	0	3	4	29.54
Ethylbenzene	219	Hydroxylation	[62]	106.16	3.21	3.21	3.21	0	0	1	0
Verbenone	91	Hydroxylation	[63]	150.22	1.97	1.97	1.97	0	1	0	17.07
Efavirenz	12.4	Hydroxylation	[6]	315.67	4.84	4.84	4.84	1	3	1	29.54
Carbamazepine	420	Hydroxylation (2 sites)	[64]	236.27	2.67	2.67	2.67	2	3	0	23.55
Carbaryl	11	Hydroxylation (3 sites)	[65]	201.22	2.40	2.4	2.4	1	3	2	29.54
(S)-Ketamine	44.4	N-demethylation	[40]	237.72	2.18	1.18	2.13	1	2	2	20.31
(R)-Ketamine	74.1	N-demethylation	[40]	237.72	2.18	1.18	2.13	1	2	2	20.31
LAAM	17.7	N-demethylation	[66]	353.50	5.45	2.42	3.48	0	3	9	29.54
Nor-LAAM	26	N-demethylation	[66]	339.47	3.76	0.67	1.11	1	3	9	29.54
S-mephobarbital	264	N-demethylation	[67]	246.26	1.77	1.77	1.67	1	5	2	57.69
Selegiline	25	N-demethylation and depropargylation	[68]	187.28	2.95	0.95	2.58	0	1	4	3.24
(S)-ifosfamide	900	Dechloroethylation	[38]	261.09	0.23	0.23	0.23	1	4	4	42.59
(R)-ifosfamide	500	Dechloroethylation	[38]	261.09	0.23	0.23	0.23	1	4	4	42.59
Cyclophosphamide	1400	Hydroxylation	[69]	261.09	0.23	0.23	0.23	1	4	5	42.59
(R)-fluoxetine	125	N-demethylation	[39]	309.33	4.09	1	1.56	1	2	6	12.47
(S)-fluoxetine	546	N-demethylation	[39]	309.33	4.09	1	1.56	1	2	6	12.47
Propofol	10	Hydroxylation	[70]	178.27	4.16	4.16	4.16	1	1	3	9.23
			Mean	253.78	2.68	1.51a	2.02b	0.57c	2.57	3.50	27.63
			(±SD)	(74.03)	(1.51)	(1.43)	(1.25)	(0.57)	(1.37)	(2.71)	(19.42)
			Mean #	249.51	2.68	1.60	2.09	0.54	2.58	3.29	28.25
			(±SD)	(78.31)	(1.49)	(1.51)	(1.29)	(0.59)	(1.44)	(2.77)	(20.45)

MW = molecular weight, HBA = hydrogen bond acceptors, HBD = hydrogen bond donors, RBN = rotatable bond number, PSA = polar surface area

^#after removal of one isomer for compounds fluoxetine, ifosphamide, ketamine and methadone

^ap < 0.009 t-test (2 tailed) compared with Table 3

^bp < 0.003 t-test (2 tailed) compared with Table 3

^cp < 0.005 t-test (2 tailed) compared with Table 3

Table 3.

CYP2B6 substrates from the original pharmacophore and test set [5] with molecular properties calculated with ChemSpider.

Substrate	Reaction	Km (µM)	MW	ACD/logP	LogDpH5.5	LogDpH7.4	HBD	HBA	RBN	PSA
Antipyrine	hydroxylation	17700	188.23	0.27	0.27	0.27	0	3	1	23.55
Benzphetamine	N-demethylation	93.4	239.35	4.43	1.51	2.95	0	1	5	3.24
Benzyloxyresorufin	N-demethylation	1.28	303.31	3.15	3.15	3.15	0	4	3	47.89
Cinnarizine	hydroxylation	17.2	368.51	4.63	3.47	4.56	0	2	6	6.48
4-chloro-7-ethoxycoumarin	O-deethylation	33.7	224.64	2.50	2.37	2.37	0	3	2	35.53
3-cyano-7-ethoxycoumarin	O-deethylation	71.3	215.20	2.00	2.51	2.51	0	4	2	59.32
Dextromethorphan	N-demethylation	350	271.40	4.11	1.12	2.39	0	2	1	12.47
Diazepam	N-demethylation	113	284.74	2.91	2.91	2.91	0	3	1	32.67
1,2-Dibromoethane	2-bromoacetaldehyde formation	9700	187.86	2.08	2.08	2.08	0	0	1	0
7-ethoxycoumarin	O-deethylation	115	190.19	2.31	2.31	2.31	0	3	2	35.53
7-ethoxy-4-trifluoromethylcou marin	O-deethylation	1.7	258.19	3.41	3.42	3.42	0	3	2	35.53
Imipramine	N-demethylation	383	280.41	4.80	1.76	2.79	0	2	4	6.48
Midazolam	Hydroxylation	46.1	325.77	3.93	3.48	3.92	0	3	1	30.18
RP73401	Hydroxylation	22.5	381.25	4.27	4.27	4.27	1	5	5	51.66
S-mephenytoin	N-demethylation	564	218.25	1.40	1.4	1.32	1	4	2	40.62
Testosterone	Hydroxylation	50.5	288.42	3.47	3.48	3.48	1	2	1	26.3
Amitriptyline	N-demethylation	144.4	277.40	4.92	1.92	3.15	0	1	3	3.24
Bupropion	Hydroxylation	107.5	239.74	3.47	1.82	3.27	1	2	4	20.31
Lidocaine	N-deethylation	537.6	234.34	2.36	−0.13	1.57	1	3	5	23.55
Verapamil	O-demethylation Dihydroqinghaosu	137.4	454.60	3.90	0.95	2.33	0	6	13	63.95
Arteether	formation	28	312.40	3.60	3.6	3.6	0	5	2	46.15
		Mean	273.53	3.23	2.27	2.79	0.24	2.90	3.14	28.79
		(±SD)	(68.35)	(1.20)	(1.17)	(1.00)	(0.44)	(1.44)	(2.76)	(18.95)

MW = molecular weight, HBA = hydrogen bond acceptors, HBD = hydrogen bond donors, RBN = rotatable bond number, PSA = polar surface area

For references for the original K_m data please refer to [5].

Fig. 1.

Sequence alignment of CYP2B6 and CYP2B4 (assession numbers P20813 and P00178, respectively) using ClustalW (EMBL-EBI) [84] * = identical in all sequences in the alignment; : = conserved substitutions; . = semi-conserved substitutions.

Since the publication of the original pharmacophore in 1999 [5, 20] a substantial number of studies have indicated the involvement of CYP2B6 in the metabolism of molecules under study. These molecules cover an array of therapeutic classes including anticancer agents, antiviral drugs, antidepressants, and pesticides [21] (Table 1–Table 2). The current study was performed to obtain a more complete picture of the characteristics of CYP2B6 substrates. These substrates themselves can provide insights into the requirements for the enzyme.

Table 2.

Additional CYP2B6 substrates without K_m data and their molecular properties calculated with ChemSpider.

Substrate	Reaction	Reference	MW	ACD/logP	LogDpH5.5	LogDpH7.4	HBD	HBA	RBN	PSA
Perhexiline	Hydroxylation	[71]	277.49	6.95	3.85	3.93	1	1	4	3.24
Bergamottin	Hydroxylation and epoxidation	[72]	338.40	5.92	5	5	0	4	6	48.67
Delta-3-carene	Hydroxylation	[73]	136.23	4.37	4.37	4.37	0	0	0	0
Chlorpyrifos	Desulphuration	[74]	350.59	4.77	4.77	4.77	0	4	6	82.48
Diazinon	Desulphuration	[74]	304.34	3.81	3.81	3.81	0	5	7	95.37
Serotonin	Deamination	[31]	176.21	0.21	−2.87	−2.22	4	3	4	17.4
Methoxychlor	O-demethylation and Hydroxylation	[75]	345.65	4.56	4.56	4.56	0	2	4	18.46
Tamoxifen	Hydroxylation	[76]	371.51	7.88	5.04	6.58	0	2	8	12.47
ThioTEPA	Desulfuration	[77]	189.22	0.52	0.52	0.52	0	3	3	50.93
Tramadol	O-demethylation and N-demethylation	[78]	263.37	2.51	−0.53	0.49	1	3	5	21.7
MDE	N-dealkylation	[79]	283.36	2.73	0.46	1.8	1	3	5	21.7
Artemisinin	Unknown	[80]	282.33	2.27	2.27	2.27	0	5	0	53.99
DPC963	Hydroxylation	[81]	316.23	4.09	3.38	3.21	2	3	1	41.13
Rhazinilam	Hydroxylation	[82]	294.39	5.43	5.02	5.02	1	3	1	25.24
Mexiletine	N-oxidation	[83]	179.26	2.16	−0.63	0.96	2	2	4	12.47
		Mean	273.91	3.88	2.60	3.00	0.80	2.87	3.87	33.68
		(±SD)	(72.09)	(2.19)	(2.56)	(2.32)	(1.14)	(1.35)	(2.47)	(28.02)

MW = molecular weight, HBA = hydrogen bond acceptors, HBD = hydrogen bond donors, RBN = rotatable bond number, PSA = polar surface area

In the current study we have collated 43 substrates from the literature, 28 with experimentally determined K_m data using recombinant expressed CYP2B6 (Table 1) and a further set of 15 without K_m data (Table 2). From analysis of the molecular properties of these additional substrates we are able to compare them with those from the original pharmacophore study [5]. In addition we have evaluated several other 2D and 3D-QSAR methods in an attempt to generate new pharmacophores and models for the CYP2B6. These would potentially be useful in enabling selection of molecules for in vitro testing that could be developed as CYP2B6 probe substrates, or, alternatively, assist in avoiding this enzyme. This work builds on our research combining computational and in vitro data analysis to improve the understanding of key molecular properties required by these cytochromes P450 [22] using methods described in an earlier review in this journal [23]. These computational analyses will supplement future efforts in further understanding the role of this enzyme.

MATERIALS AND METHODS

Descriptor data and statistical analysis

A comprehensive literature search was performed to identify molecules with K_m data generated with recombinant expressed CYP2B6 that had not been previously included in our prior CYP2B6 QSAR model [5]. In most cases we are aware this data is generated with B-lymphoblastoid cells [5] or supersomes expressing CYP2B6, however, rarely is it mentioned whether cytochrome b₅ was added or how much P450 reductase was co-expressed. It was therefore difficult to ensure that all papers used exactly the same experimental reagents and protocol. 2D structures of these 43 molecules (28 substrates with K_m data, 15 substrates without K_m data) were taken from the ChemSpider database (www.chemspider.com). Molecular weight (MW), ACD/LogP, LogD _{pH 5.5}, LogD _{pH 7.4}, HBA, hydrogen bond donor (HBD), polar surface area (PSA), rotatable bond number and molecular weight (MW) were also obtained from the ChemSpider database. The octanol-water partition coefficient (LogP, also referred to as K_ow) is a measure of the propensity of a neutral compound to differentially dissolve in these immiscible phases. The octanol-water distribution coefficient, LogD, reflects the true behavior of an ionizable compound in a solution at a given pH value or range, taking into account all the different ionized species present. This molecular property generation approach was repeated for the 21 molecules from the original CYP2B6 pharmacophore study [5]. The distribution of molecular property data was statistically analyzed with SPSS 14.0 (SPSS Inc. Chicago IL) and mean (±SD) values for molecular descriptors compared with the mean (±SD) data for the 21 molecules using the t-test (2-tailed).

In silico modeling: 2D-QSAR analysis

Forty molecular descriptors were calculated using the QuASAR module in MOE (Chemical Computing Group, Montreal, Canada) for the training and test sets to make up the initial descriptor “pool”. Selection of the most relevant descriptors was achieved by using QuaSAR-Contingency which performs a bivariate contingency analysis for each descriptor and the activity value. Four molecular descriptors with top contingency coefficients were selected to construct a linear model to fit the CYP2B6 data. The training set consisted of 21 CYP2B6 substrates from the original pharmacophore set (Table 3). Partial least squares (PLS) regression was used to determine the parameters of the linear model with all entries being equally weighted.

In silico modeling: Catalyst™

Computational molecular modeling studies were carried out using Catalyst version 4.9 in Discovery Studio version 1.7 (Accelrys, San Diego, CA) running on a Sony Vaio computer with a Pentium processor. Pharmacophore models attempt to describe the arrangement of key features that are important for biological activity. Briefly, the previously generated CYP2B6 Catalyst substrate model [5] was employed to map new molecules after first importing the molecule structures as an sdf file in Catalyst. The 3-D molecular structures were produced using up to 255 conformers with the Best conformer generation method, allowing a maximum energy difference of 20 kcal/mol. A shape was generated around the 7-ethoxy-4-trifluoromethylcoumarin (7-EFC) structure mapped to the original pharmacophore and this was used for further mapping attempts with test molecules.

In silico modeling: Phase™

Computational molecular modeling was also performed with Phase, version 2.0 (Schrödinger, Portland, OR). Activities for all compounds were defined as pK_m, where p = −log is in molar units. Phase is a second pharmacophore elucidation and QSAR model building method [24, 25]. Molecules used in the training and test set were first processed in LigPrep, with the protonation states appropriate to pH 7 assigned. Multiple conformers for each molecule were then generated in MacroModel through systematic sampling around rotatable bonds using the UPLS_2005 force field. The default pharmacophore feature definitions were used for assignment of pharmacophore points. Hypotheses were generated by a systematic variation of the number of sites and number of matching active compounds. The scoring in Phase was performed using the default parameters, and the score weighted to include alignment of the inactive compounds also used the default parameters. All hypotheses that were successfully generated and scored were then used to build 3D-QSAR models with 1–3 PLS factors. Statistics on the correlation of predicted versus actual activity were collated for the top ten scoring hypotheses utilizing the default hypothesis scoring functions.

RESULTS

Descriptor data and statistical analysis

A thorough analysis of the literature enabled us to accrue a new database of 28 substrates with K_m data for recombinant expressed CYP2B6. Following simple statistical analysis of the data we have observed that these molecules tend to be small (mean MW 253.78±74.03) and hydrophobic (mean ACD/logP 2.68 ±1.51, LogD _{pH 5.5} is 1.51±1.43, LogD _{pH 7.4} is 2.02±1.25). The HBA count is 2.57±1.37, the number of rotatable bonds is 3.50±2.71 and the PSA is 27.63±19.42 (Table 1). CYP2B6 substrates also generally appear to be devoid of HBD features with 0.57±0.57 (Table 1). These mean values did not change dramatically upon considering just a single isomer of the four compounds included in the analysis (Table 1), or after focusing only on molecules with K_m < 50µM (data not shown). Lower K_m molecules tend to be slightly larger with a slightly increased ACD/logP; however there was only a weak correlation between ACD/logP and −logK_m (r² = 0.24).

A previously devised QSAR relationship for 16 CYP2B6 substrates [12] used the following equation using Log P generated with Pallas Software System (CompuDrug Ltd., Budapest, Hungary):

$$-\text{log}{\mathrm{K}}_{\mathrm{m}}=0.88\text{log}\mathrm{P}+1.68$$ equation 1

When equation 1 was used to calculate the −logK_m for the 35 substrates not used to build equation 1 and compared with the experimental determined −logK_m values the correlation was low (r² = 0.22). The latter observations suggest that this equation may need updating and that it may be difficult to predict an exact K_m value using a single property like LogP alone and more complex models may be required for CYP2B6 substrates.

A second set of 15 molecules without published CYP2B6 K_m data (to our knowledge) had very similar mean molecular property values to the initial 28 compound set with MW = 273.91±72.09, ACD/logP = 3.88±2.19, LogD _pH5.5= 2.60±2.56, LogD _pH7.4 = 3.00±2.32, HBD = 0.80±1.14, HBA = 2.87±1.35, rotatable bond number 3.87±2.47 and PSA 33.68±28.02 (Table 2). These data are also remarkably similar to those obtained with the original 21 molecules used in the pharmacophore study [5] in which MW = 273.53±68.35 , ACD/logP = 3.23±1.20, LogD _pH5.5 = 2.27±1.17, LogD _pH7.4 = 2.79±1.00, HBD = 0.24±0.44, HBA = 2.90±1.44, rotatable bond number 3.14±2.76 and PSA 28.79±18.95 (Table 3). Following t-test analysis comparing the mean values of this original set with the separate sets of 28 and 15 substrates, only the HBD count and the logD values were significantly different for the 28 substrates. All other comparisons did not represent statistically significant differences. Incidentally in all three data sets examined the HBD count mean value is consistently less than 1 and the variability is large, indicating that this molecular property is unlikely to be important for CYP2B6 substrates and this feature may frequently be absent in substrates. Overall statistics (with standard deviations in parentheses) for all the 64 substrates assessed in this study are MW = 264.98 (71.31), ACD/logP = 3.14 (1.65), LogD _pH5.5 = 2.01 (1.73), LogD _pH7.4 = 2.51 (1.54), HBD = 0.52 (0.73), HBA = 2.75 (1.38), rotatable bond number = 3.47 (2.65), PSA = 29.43 (21.35) and these are similar to the corresponding means in each of the three subsets described above (Fig. 2). The histograms for the data generally show normal distributions for the molecular properties apart from HBD, PSA and RBN which appear skewed or bimodal (Fig. 2).

Fig. 2.

Histograms of the molecular descriptor distributions of all 64 CYP2B6 substrates used in this study. A. Molecular weight (MWT), B. ACD/logP, C. LogD pH5.5, D. LogD pH 7.4, E.hydrogen bond donors (HBD), F. hydrogen bond acceptors (HBA), G. rotatable bond number (RB), H. polar surface area (PSA).

In silico modeling: 2D-QSAR analysis

The QuaSAR module in MOE was used to perform the 2D-QSAR PLS regression analysis on the data set. Four molecular descriptors with top contingency coefficients were picked that fit to the CYP2B6 data. The training set consisted of 21 CYP2B6 substrates from the original pharmacophore set (Table 3). This resulted in a model with the following statistics: N=21, r² = 0.82, cross validated (leave one out) r² = 0.69:

$${\text{pK}}_{\mathrm{m}}=2.11-0.015*\text{MW}+0.55*\text{log}{\mathrm{P}}_{(\mathrm{o}/\mathrm{w})}+0.025*\text{TPSA}-0.86*\text{log}\mathrm{S}$$ equation 2

Where MW is Molecular weight; logP_(o/w) is Octanol/Water partition coefficient; TPSA is Topological Polar Surface Area; logS is Log Solubility in water. The normalized linear model showed the relative importance of the four descriptors in the model to be 1.00 for Weight, 0.94 for logS, 0.65 for logP_(o/w) and 0.49 for TPSA. In this linear model, 7-EFC was an outlier, which when omitted improved the model [r² to 0.88 and cross validated (leave one out) r² to 0.73] statistics slightly. Equation 2 was used to predict the 28 test set molecules (Table 4) with r²= 0.396, but significant rank ordering, Spearman’s rho = 0.663 (p = 0.01, two tailed). It should be noted that while all 28 molecules were predicted, cyclophosphamide and its isomers S- and R-Ifosfamide were poorly predicted with the same value (−0.155). This may be related to the calculated LogP (Table 1) or Log S values. One of the limitations using the descriptors in this model is the inability to discriminate between isomers.

Table 4.

Test set prediction using 2D-QSAR model equation 2

Name	Experimental pKm (M)	Predicted pKm (M)
Nevirapine	3.079	2.609
R-Methadone	4.387	4.251
S-Methadone	4.678	4.251
N-methyl,N-propargyl-2-phenylethylamine	3.538	2.826
(+)-Fenchone	3.745	2.553
Endosulfan Alpha	4.790	4.910
Prasugrel	5.638	3.706
NNK	4.481	1.465
S-nicotine	3.086	1.325
Sertraline	4.513	5.277
Meperidine	3.449	2.862
Ethylbenzene	3.660	4.235
Verbenone	4.041	3.045
Efavirenz	4.907	5.608
Carbamazepine	3.377	4.709
Carbaryl	4.959	4.536
R-Ketamine	4.130	3.427
S-Ketamine	4.353	3.427
LAAM	4.752	4.622
nor-LAAM	4.585	4.793
S-Mephobarbital	3.578	3.616
Selegiline	4.602	3.153
S-Ifosfamide	3.046	−0.155
R-Ifosfamide	3.301	−0.155
Cyclophosphamide	2.854	−0.155
R-Fluoxetine	3.903	4.291
S-Fluoxetine	3.263	4.291
Propofol	5.000	5.030

In silico modeling: Catalyst™

The coordinates of the previously published CYP2B6 pharmacophore have already been described and initially validated with a small test set [5]. Nineteen of the 28 (68%) test set molecules in this study were predicted within a 1 log residual using the rigid fit as performed in the original publication (Table 5). The addition of the 7-EFC shape to the pharmacophore to create a shape/feature hypothesis which was used to classify molecules, retrieved 16 of the 28 molecules (of which 11 were within the 1 log residual (68%)). Therefore this method of filtering the molecules did not appear to improve the percentage within 1 log residual. Also, there was no significant correlation between the log observed and predicted K_m data for the test set which indicates that the pharmacophore needs updating in order to quantitatively predict these new molecules. Attempts to do this to date have been unsuccessful. This could be attributable to the different conformer generation method now implemented in Discovery Studio Catalyst compared to the previous version 3.1 used in the original paper, the training set may now be too large or more likely, the data in the training set may represent molecules binding in multiple locations, creating noise which could be detrimental to the pharmacophore construction. This may also make it difficult to ensure predictions with much lower residuals than 1 which would be preferable and more valuable to drug discovery.

Table 5.

Predictions with the original CYP2B6 pharmacophore [5].

Molecule	Observed Km (µM)	Rigid Fit predicted Km (µM)	Log residual	Shape Mapping
Nevirapine	834	129.3	−0.81	y
(R)- methadone	41	5.3	−0.89
(S)- methadone	21	32.7	0.19
N-methyl,N-propargyl-2-phenylethylamine	290	7762.5	1.43
(+)-fenchone	180	295.9	0.22
Endosulfan Alpha	16.2	30	0.27
Prasugrel	2.3	31.2	1.13
NNK	33	30.7	−0.03	y
(S)-Nicotine	820	150.8	−0.73	y
Sertraline	30.7	50.2	0.21	y
Meperidine	356	6.4	−1.7	y
Ethylbenzene	219	9045.5	1.61
Verbenone	91	290.84	0.50
Efavirenz	12.4	80.4	0.81
Carbamazepine	420	102.4	−0.61	y
Carbaryl	11	45.1	0.61	y
(S)-Ketamine	44.4	35.8	−0.09	y
(R)-Ketamine	74.1	34.2	−0.33	y
LAAM	17.7	1.7	−1.02
Nor-LAAM	26	56.9	0.34
S-mephobarbital	264	167.5	−0.20
Selegiline	25	28.3	0.05	y
(S)-ifosfamide	900	1.2	−2.87	y
(R)-ifosfamide	500	0.4	−3.10	y
Cyclophosphamide	1400	36.2	−1.59	y
(R)-fluoxetine	125	81.3	−0.19	y
(S)-fluoxetine	546	83.4	−0.82	y
Propofol	10	561.9	1.75	y

y = Yes

In silico modeling: Phase™

The Phase 3D-QSAR method is used to search for a common pharmacophore (Similar to the Catalyst HIPHOP approach [26]) within a small set of "actives" and then the remaining molecules should align to this hypothesis for developing the QSAR model. Starting with all the training molecules we were able to obtain a 4-point hypothesis for 6 of the 21 molecules in Table 3. Phase was only able to align 14 of the 21 to the best hypothesis (i.e. only 14 fit at least 3 of the 4 points). Benzphetamine, cinnarizine, dibromoethane, imipramine, testosterone, amitriptyline and bupropion could not be aligned to this pharmacophore. The 14 molecule training set was able to generate a reliable QSAR model with r²= 0.69. This pharmacophore consisted of two HBAs, a hydrophobic feature and an aromatic ring (a second hydrophobic feature). However, the Phase pharmacophore itself was only able to make predictions for 15 of the 28 molecules in Table 1, and the observed versus predicted log K_m was not correlated significantly (data not shown).

DISCUSSION

A previous study has suggested that hydrophobicity as commonly measured by calculated logP may be a good predictor for affinity to CYP2B6 [12]. In the current study we found only a weak correlation between predicted −logK_m and observed −logK_m using the suggested simple QSAR (equation 1). In addition we observed similar weak correlations when relating logP and −logK_m, therefore indicating that more complex methods may be necessary to accurately predict affinity for this enzyme, which is generally in line with the models that have been developed for other P450s [27]. Analysis of several other calculated molecular properties did not indicate one with a single correlation that improved upon calculated logP. The ChemSpider database and molecular descriptor calculator used in this study is a freely available tool utilizing widely used and validated property calculators [28]. Other freely available sources of the descriptors used in this study such as XlogP from PubChem and ACD/logP used in this study, had a good correlation with each other (r²= 0.79, data not shown) across all 64 molecules. Similarly, a comparison of all molecular properties calculated by both PubChem and ChemSpider indicated there was no significance difference between them (data not shown). Analysis of the molecular properties for these 64 molecules indicates that overall CYP2B6 is quite selective in the molecular requirements for substrates. For example the majority (~80%) of the 64 CYP2B6 substrates used here are within the molecular weight range of 200–350 and most molecular properties apart from HBD and TPSA and RBN appear to be normally distributed (Fig. 2). Using a 2D-QSAR method with the 21 training set molecules and just 4 molecular descriptors, MW, LogP_(o/w), LogS and TPSA, we were able to improve the correlation (r² = 0.82, cross validated (leave one out) r² = 0.69) as well as being able to statistically significantly rank order a test set of 28 molecules. This represents an improvement over an earlier 2D-QSAR model for 16 molecules (equation 1, [12]).

The low PSA of the molecules would suggest most CYP2B6 substrates could cross the Blood Brain Barrier (BBB) [29] and have central nervous system (CNS) effects, though this may be counteracted by potential P-gp affinity of these relatively hydrophobic molecules and both of these properties could be predicted computationally [30]. From the overwhelming weight of evidence now presented, CYP2B6 may therefore have an important role in the metabolism of many CNS active molecules e.g. serotonin [31], methadone, sertraline, meperidine, propofol, mephenytoin, carbamazepine, burproprion, ketamine, nicotine etc., as well as compounds with undesirable CNS effects such as ifosfamide (Table 1). This may represent an area of particular interest to the pharmaceutical industry. A previous study had analyzed various CNS acting classes of drugs in the Comprehensive Medicinal Chemistry (CMC) database and suggested a difference between mean log P for antipsychotic (4.1), antidepressant (3.1) and hypnotic (2.2) groups with a standard deviation of 1.5 [32]. The current study determined a mean ACD/LogP for the 64 molecules of 3.14, which would indicate that antidepressants and hypnotics are most likely to be substrates for CYP2B6. Other members of these classes of compounds could be prioritized for testing using the molecular property criteria developed in this study. This may be clinically relevant as CYP2B6 is expressed in the brain and is inducible [33]. The majority of the 64 metabolic reactions catalyzed were N-demethylation and hydroxylations, although several other reactions were less frequently observed. This information could additionally be useful in generating a probability for CYP2B6 metabolizing a molecule at these corresponding functional groups [34] or building models to predict rate of metabolism [35]. Ultimately models for V_max, clearance and/or K_cat for this enzyme would also be useful.

As we have previously generated a pharmacophore for 16 CYP2B6 substrates we assessed the mapping of the 28 molecules to the pharmacophore as a further large test set. Although it should be noted that no significant correlation was observed, 68% of molecules were predicted within 1 log unit which was an acceptable (though not ideal) criteria used in the original study. This percentage did not improve with the inclusion of the 7-EFC shape in this study, which would provide a stringent molecular shape/volume for other substrates to fit into (Fig 3A, B). Interestingly all 16 molecules selected fit well within this shape volume, and this may indicate just one potential pharmacophore for these molecules with CYP2B6 as indicated previously using a slightly different training set [11]. The remaining molecules likely fit to another pharmacophore which may correspond with those developed by Wang and Halpert [36]. From the literature analysis we have also observed that several studies have now indicated that CYP2B6 demonstrates stereoselectivity, namely for methadone [37], ifosfamide [38], fluoxetine [39] and ketamine [40] (Table 1). This is in addition to the well known stereoselectivity of S-mephenytoin by CYP2B6 [41]. This would suggest perhaps the further investigation of the stereoisomers of other racemic CYP2B6 substrates as stereoselectivity could be an important component of this enzyme. A recent study with individual isomers of bupropion [42], (previously identified as a CYP2B6 substrate and now a widely used probe reaction [5, 43, 44]) indicated stereoselectivity [K_m (R) vs (S), 46µM and 34 µM respectively]. When compared to the racemate with a K_m of 107.5µM [45]. The difference between the isomers could be due to the positioning of the terminal methyl which may interfere with the hydrogen bonding interaction at the C=O in R-bupropion, while it is not in this position in S-bupropion. The pharmacophore failed to predict the correct order of the R- and S- methadone and ketamine isomers, but it rank ordered the R- and S- ifosfamide and fluoxetine isomers. This may indicate the need for excluded volumes to more accurately represent the binding site of CYP2B6 possibly guided by the homology model, in order to further improve the predictions. An initial homology model for CYP2B6 did not account for stereoselectivity [36], while a second model took regioselectivity into account for cyclophosphamide and ifosfamide [46] as three pockets were required to be filled for observation of product formation. Our attempts to update the Catalyst models with new data and utilize excluded volumes have been unsuccessful to this point and this could be due to data variability or the inclusion of stereoisomers. We also attempted to use additional 3D-QSAR methods including GALAHAD [47] and a 4D-QSAR method [48] but were also unable to generate a significant model with these software tools, which further indicated the complexity perhaps inherent in this dataset (stereoselectivity and multiple pharmacophores). Our attempts with the pharmacophore software, Phase was partially successful, producing a pharmacophore with 2 hydrogen bonds and 2 hydrophobic areas, using 14 of the 21 training molecules (Fig 3C). This pharmacophore partially overlaps with that of the Catalyst model (Fig 3A, B). However, this Phase model was unable to generate reliable predictions for an external test set measured in terms of a correlation with published observed K_m data. Overall, our application of several commercially available 3D-QSAR methods may indicate that this available CYP2B6 K_m data is complex and difficult to obtain predictive models with, this may also be partially attributable to the data being derived from different laboratories, using a variety of recombinant expression systems and incubation conditions, factors that may affect observed kinetics in vitro.

Fig. 3.

A. NNK mapped to the original CYP2B6 pharmacophore, showing shape volume from 7-EFC. Note N-methyl group on left side is site of α-hydroxylation. H = hydrophobe, NBA = hydrogen bond acceptor B. Mapping of 16 of 28 substrates that conform to the 7-EFC shape. Pharmacophore shows 3 hydrophobic features (H) and one hydrogen bond acceptor (HBA) with vector. C. Phase pharmacophore with benzyloxyresorufin, A = hydrogen bond acceptor (with arrows), H = hydrophobe (sphere), R = ring aromatic (circle).

Interestingly, previously allosteric kinetics were also noted for methadone N-demethylation [37] while several other studies had indicated similar allosteric or autoactivation kinetics with 7-EFC [2], testosterone 16β-hydroxylation [3] and verapamil [4, 5, 49]. From our analysis of the molecules included in Table 1–Table 3, we observed published studies do not generate full kinetic curves (to derive a K_m and V_max) with recombinant expressed CYP2B6, that in turn have significant numbers of observations at the low concentrations in order to pick up the “hook” in an Eadie-Hofstee plot. Sometimes an apparent “hook” on the Eadie-Hofstee plot is missed altogether. Others have demonstrated the role of CYP2B6 in the 8-hydroxylation of efavirenz (Table 1) and noted allosteric kinetics in human liver microsomes (HLM) but not in recombinant expressed CYP2B6 [6]. However, on closer inspection they used more data points at lower concentrations in HLM (14 versus 10 points overall). So it would appear that allosteric characteristics are possibly under-reported for CYP2B6 substrates and this behavior could be reserved for certain molecular structures, or may be due to experimental conditions in vitro. Allosteric kinetics may also be a contributing factor in the difficulty in generating predictive 3D-QSAR models.

Based on the published in vitro and computational models it was previously suggested that the binding site for CYP2B6 may enable either multiple binding modes for a substrate, distinct binding sites or there is substantial conformational change in the binding site upon ligand binding [36, 45, 46], these would all make it quite difficult to come up with a single general model for this enzyme [3] as we have demonstrated in this study. Interestingly, Wang and Halpert developed 2 pharmacophore models that suggested the possibility of 2 overlapping sites, although this was not explicitly mentioned in their discussion [11]. Bathelt et al., also indicated three hydrophobic pockets that may exist in the CYP2B6 homology model binding site [46]. The original CYP2B6 pharmacophore provided a similar conclusion in a hydrophobic binding site [5] (one of the hydrophobic features aligns with the site of metabolism feature [11]) and is almost identical to that of the CYP3A4 autoactivator pharmacophore [50], both containing 3 hydrophobes and a central HBA feature. This in many ways mirrors the observation that some substrates are CYP2B6 or CYP3A4 selective while others have overlapping activity with both enzymes. It would also appear that CYP2B6 is less promiscuous than CYP3A4 [51] based on the narrow range observed for molecular properties determined in this study. In comparison, CYP3A4 substrates range widely in size and other properties with molecules (K_m < 10µM) possessing high LogP, RBN [52] compared with weaker substrates (K_m > 100µM) having LogP and RBN closer to the CYP2B6 mean values in this current study. This also indicates overlap in the requirements (pharmacophores) for both enzymes which are fulfilled by some substrates (co-substrates) and not others (selective substrates). It could be useful to differentiate these molecules separately (including those from this literature analysis) and use them in future model building. Also, the published CYP2B6 pharmacophore [5] is generally similar to a previously published constitutive androstane receptor (CAR) pharmacophore, suggesting CYP2B6 substrates would appear to have features in common with the ligands that activate the major nuclear hormone receptor that transcriptionally regulates CYP2B6 gene expression [53]. Similarly, the CAR pharmacophore although smaller, partially overlaps with the PXR pharmacophore [53], in agreement with the experimental data with known inducers [18].

The current QSAR and statistical analysis using simple, interpretable molecular properties that are freely available from ChemSpider, PubChem (or other reliable sources of chemistry data on the internet), suggests that 64 molecules (Table 1, Table 2 and Table 3) demonstrating CYP2B6 affinity are to varying degrees hydrophobic, with one or more HBAs, and in good agreement with the properties obtained in previous pharmacophore and QSAR models for CYP2B6. These properties may overlap with those of low affinity CYP3A4 substrates. Attempts to generate a predictive 2D-QSAR model appeared to be more successful than the 3D-QSAR pharmacophores as the 2D-QSAR model was able to rank order the test set molecules. Further computational studies will likely be pursued to construct updated models for CYP2B6 substrates and these findings will be of value for pharmaceutical research to enable prediction of molecule-CYP2B6 interactions from molecular structure alone. The availability of crystal structures for the Rabbit CYP2B4 [9, 10] will also be useful for comparisons with CYP2B6 homology models as well as updating them due to the increased sequence identity (Fig 1). While substrate selectivities for CYP2B6 and CYP2E1 have been compared, to our knowledge we are not aware of a similar comparison between CYP2B6 and CYP2B4 [16]. The generation of a large number of CYP2B6 K_m values from a single laboratory and further evaluation of potential for autoactivation may go some way to offset the variability in the published data. It will also be of considerable interest to observe whether the range of property values described in this study holds for future CYP2B6 substrates to be identified.

Abbreviations

BBB

Blood Brain Barrier

CAR

constitutive androstane receptor

PXR

pregnane X-receptor

QSAR

quantitative structure activity relationship