Predictive Model for Identifying Potential CYP2D6 Inhibitors
Patrizia Crivori, Italo Poggesi · Basic & Clinical Pharmacology & Toxicology · 2005
The inhibition of cytochrome P450 2D6 (CYP2D6) by drug candidates is a liability that discovery programmes try to avoid (Steiner et al. 1998). Computational models for identifying inhibitors of this isoform would be of great utility for helping the correct selection of compound devoid of such issue. Some models for identifying CYP2D6 substrates (Islam et al. 1991; de Groot et al. 1997) and inhibitors (Strobl et al. 1993; Ekins et al. 1999) already appeared in the literature. The purpose of this study was to develop and evaluate a model for predicting the CYP2D6 inhibition potential based on the molecular structures of drug candidates. The identification of common pharmacophoric features of the CYP2D6 inhibitors under analysis was also an objective of this evaluation. The apparent inhibition constants (Ki) for 64 compounds, determined in vitro using human liver microsomes, were collected from the literature (Strobl et al. 1993; Ekins et al. 1999). The range of apparent Ki was 0.0046 to >1000 μM. The inhibition data were divided into a subset of 47 compounds used for training the QSPR model, and a subset of 17 compounds to validate it (Test set 1). An additional set of 33 apparent Ki, was taken from internal discovery projects to further evaluate the model (Test set 2). The structures of the inhibitors were built up interactively using Cerius2 program (Accelrys, v. 4.9) and optimized using MERK Force Field (MMFF). Conformational analysis was carried out using Catalyst (Accelrys, v. 4.7). The structures of the test sets were built up interactively using Cerius2 program or retrieved from internal database. A single conformation was generated using CORINA (Molecular Networks, v. 2.6) and minimized by MMFF force field. Grid-independent descriptor (GRIND) generation involved various steps, performed automatically by the program Almond (Molecular Discovery, v. 3.2a). The details of this procedure is reported by Pastor et al. (2000) and Fontaine et al. (2004). In our study, the molecular interaction fields for three probes (DRY, N1 and CO) and TIP “shape-field” were generated. All molecular interaction fields were computed using a grid spacing of 0.5 Å with the grid extending 5 Å beyond the molecule. The maximum number of extracted nodes was set to 150 and molecular interaction field weights were decreased to 35%. Ten correlograms, containing 61 variables each, were calculated. The QSPR model was developed by correlating the experimental inhibition constants (expressed as pKi) with the unscaled descriptors calculated for all conformers (140) generated for the training set. For this purpose, partial least squares (PLS) analysis was performed using Almond. To avoid overfitting only the first two latent variables were used to predict the two external test sets. The correlation coefficient of the model r2 was 0.62 and the correlation coefficient q2 after three random groups cross-validation was 0.50. The inhibition constants of 17 known drugs and of 33 proprietary compounds were predicted using the QSPR model. The experimental versus predicted Ki (μM) are shown in fig. 1. Thirty-eight % of the entire test set was predicted with a Ki error 1 log unit. Considering a threshold of 5 μM Ki for classifying compounds as strong and weak inhibitors, 78% of the entire external set was correctly predicted, with only 16% false negatives. Experimental and predicted Ki (μM) (entire Test set). In the plot the identity line and the lines representing one-half and one log-factor from identity are reported. With respect to the model interpretation, the most important features that drive the CYP2D6 inhibition potential are highlighted in fig. 2. In summary, potent inhibitors are characterized by: a) presence of favorable interacting regions around two H-bond donor groups separated by a distance of ∼13 Å (O-O 13 Å) (fig. 2a), b) presence of favourable interacting regions around an aromatic ring and a H-bond donor group separated by a distance of ∼8 Å (Dry-O 8Å) (fig. 2b), c) presence of a pocket 21 Å distant from the favorable interacting region around the H-bond donor group (O-TIP 21Å) (fig. 2c), d) presence, in the enzyme active site, of two hydrophobic pockets, one of them favoring π-π interaction (Dry-TIP 27Å) (fig. 2d), separated by a distance of 34 Å (TIP-TIP 34 Å) (fig. 2d1), e) presence of a π-π interacting pocket 34 Å distant from a favorable interacting region around an H-bond acceptor group (N1–TIP 34Å) (fig. 2e). PLS coefficients of the model (upper panel) and pharmacophoric features characterizing CYP2D6 inhibitors highlighted in ajmalicine (lower panel). GRIND descriptors appear to be a valuable tool for modeling CYP2D6 inhibitors. Although the experimental data utilized to train the model comes from different labs, the GRIND-based model appears robust and useful to predict CYP2D6 inhibition constants of external sets of compounds. The performance of this model is at least as good as that reported for other computational models of CYP2D6 inhibitors reported in the literature (Ekins et al. 1999). The use of the model for classification purposes also gave good results. Since the Almond programme allows backtracking the descriptors in 3D space, the interpretation and identification of the structural features characterizing potent CYP2D6 inhibitors was straightforward. This information can be used to reduce the inhibition potency of interesting compounds by appropriate structural modifications.