Introduction

Top Abstract Introduction Experimental Procedures Results Discussion References

In recent years several drugs have been withdrawn from the market due to cardiovascular toxicity associated with QT interval prolongation. Considerable interest in predicting this effect by noncardiovascular drugs earlier in their development has occurred since the issuance by the European pharmaceutical regulatory authority, the Committee of Proprietary Medicinal Products, of a position on QT interval prolongation in 1997 (Committee of Proprietary Medicinal Products/986/96). The focus of many in vitro studies to date is the membrane-bound inward (rapid activating delayed) rectifier potassium channel (I_Kr) [also known as the product of the human ether-a-go-go-related gene (hERG)]. This channel contributes to phase 3 repolarization by opposing the depolarizing Ca²⁺ influx during the plateau phase (Crumb and Cavero, 1999). Drugs or their metabolites may block this channel, thereby prolonging the QT interval and in some cases leading to the potentially life-threatening ventricular arrhythmia. QT prolongation may frequently result in torsades de pointes (twisting of the points), which refers to the sinusoidal variation in the QRS axis around the isoelectric line of the electrocardiogram. The end result of torsades de pointes is a ventricular tachyarrhythmia, with the prolongation of the QT interval of the last sinus beat that precedes the onset of arrhythmia. Possession of a mutation in hERG (Curran et al., 1995) or KCNE2 (Sesti et al., 2000) in the form of a single nucleotide polymorphism, may make carriers particularly sensitive to xenobiotics that in turn affect potassium currents and trigger arrhythmic events (Crumb and Cavero, 1999). It would be of considerable value in drug discovery to understand the structural requirements of inhibitors of this potassium channel before significant investment is made in a clinical candidate that may ultimately prove to be a potent hERG inhibitor. The understanding of important structural features of molecules (the pharmacophore) that inhibit hERG would enable the prediction of inhibition before molecule synthesis. Such information would reduce the likelihood of developing drugs that could lead to a life-threatening ventricular arrhythmia. At present, various in vivo and in vitro models for QT prolongation and subsequent arrhythmia exist but they may not be entirely predictive for humans. Perhaps the closest model to the human in vivo situation would be healthy human-derived cardiac tissue, but this is not readily available (Rees and Curtis, 1996). However, various cell systems expressing the hERG channel have been developed using Xenopus oocytes (Sanguinetti et al., 1995) and mammalian cell lines such as human embryonic kidney (HEK)-293 (Smith et al., 1996). The latter are perhaps more amenable to higher throughput testing but are themselves beset with limitations due to the level of expression of the channel.

A recent article on the subject of QT prolongation states that "the main objection against any novel approach to determine safety is whether the proposed tests are sufficiently powerful to reveal, at least as well as established methods, a possible adverse effect of a compound" (Crumb and Cavero, 1999). With this in mind, the aim of the present study was to critically evaluate a previously generated predictive computational pharmacophore model for hERG inhibitors built using in vitro data for antipsychotic drugs (Crumb et al., 2001). This model suggested one ring aromatic and three hydrophobic features were important for antipsychotics to block potassium conductance. For this pharmacophore, the r² correlation of the log-transformed observed versus predicted IC₅₀ values was 0.77. The power of this model to predict hERG inhibition was demonstrated by its ability to correctly rank order the IC₅₀ data for olanzapine and its two structurally related metabolites. The value for such a predictive hERG channel inhibitor model or a three-dimensional quantitative structure-activity relationship (3D-QSAR) is considerable in that the present in vitro models are costly, labor-intensive, not widely available, and therefore generally only performed on promising late discovery or development compounds. The possibility of a computational hERG model to be used as a filter in the discovery process would add an extra dimension to lead optimization. A quantitative model would provide a means of rank ordering compounds, theoretically enabling virtual selection of candidates with the lowest potential to cause hERG inhibition.

The computational approach used previously and in this study is the commercially available Catalyst software. Catalyst is a 3D-QSAR technique that generates a representative set of conformers of molecules in a training set that accounts for the maximum occupation of conformational space of chemical functionalities. Catalyst, unlike another 3D-QSAR technique, comparative molecular field analysis, does not require manual alignment of molecules that would be problematic for structurally diverse molecules. Instead, Catalyst generates a model from the chemical features of the appropriate conformers of training set molecules and represents features involved in interactions with the target after correlating measured and estimated biological activity. The initial analysis of the hERG antipsychotic-derived pharmacophore was further tested with more molecules generated under the same conditions as well as hERG inhibition data for 15 molecules from the literature. In addition a second pharmacophore, a General hERG model was generated from the same literature data and then used to predict the inhibition of a test set of 22 molecules with IC₅₀ data for hERG.

	Experimental Procedures

Top Abstract Introduction Experimental Procedures Results Discussion References

Materials. Nicotine was purchased from Aldrich Chemical (Milwaukee, WI), ketoconazole was purchased from ICN Biomedicals (Cosa Mesa, CA), and all other molecules were synthesized at Lilly Research Laboratories (Indianapolis, IN) or obtained and purified from prescribed medications.

Transfection and Cell Culture. HEK-293 cells were stably transfected through the LipofectAMINE (Invitrogen, Carlsbad, CA) method with the hERG clone. Cells were maintained in minimum essential medium with Earle's salts supplemented with nonessential amino acids, sodium pyruvate, penicillin, streptomycin, and fetal bovine serum.

Solutions. Drugs were dissolved in either dimethyl sulfoxide or deionized H₂O to make 10 mM stock solutions, which were stored at 20°C. Dilutions of stock solutions were made immediately before the experiment to create the desired concentrations. The external solution (solution bathing the cell) used for recording hERG had an ionic composition of 137 mM sodium chloride, 4 mM potassium chloride, 1.8 mM calcium chloride, 1.2 mM magnesium chloride, 11 mM dextrose, 10 mM HEPES, adjusted to a pH of 7.4 with sodium hydroxide. The internal (pipette) solution had an ionic composition of 130 mM potassium chloride, 1 mM magnesium chloride, 10 mM sodium ATP, 5 mM EGTA, 5 mM HEPES, pH 7.2, using potassium hydroxide. Experiments were performed at 37 ± 1°C.

Data Acquisition and Analysis. Currents were measured using the whole-cell variant of the patch-clamp method (Crumb, 2000). Pipette tip resistance was approximately 1.0 to 2.0 M when filled with internal solutions. Analog capacity compensation and 40 to 60% series resistance compensation were used to yield voltage drops across uncompensated series resistance of less than 3 mV. Bath temperature was measured by a thermistor placed near the cell under study and was maintained by a thermoelectric device (model 806-7243-01; Cambion/Midland Ross, Cambridge, MA). An Axopatch 1-B amplifier (Axon Instruments, Union City, CA) was used for whole-cell voltage clamping. Creation of voltage-clamp pulses and data acquisition were controlled by an IBM PC running pClamp software (Axon Instruments).

Molecular Modeling with Catalyst. The computational molecular modeling studies were carried out using a Silicon Graphics Octane workstation. Briefly, models were constructed using Catalyst, version 4.5 (Molecular Simulations, San Diego, CA) as described previously (Crumb et al., 2001). Catalyst models were also constructed with in vitro literature IC₅₀ values derived from cells expressing hERG (Table 1). Catalyst automatically uses a log transformation on these data. The number of conformers generated using the best functionality of the program for each inhibitor was limited to a maximum number of 255, with an energy range of 20 kcal/mol. Hydrophobic, ring aromatic, hydrogen bond donors, hydrogen bond acceptors, and positive ionizable features were selected for possible inclusion. Ten hypotheses were generated using these conformers for each of the molecules and the IC₅₀ values. After assessing all 10 hypotheses generated, the lowest energy cost hypothesis was considered the best because this hypothesis possessed features representative of all the hypotheses. The reliability of the structure-activity correlation between the log-transformed predicted and observed activity values was estimated by means of an r² value.

Validation of Catalyst hERG Channel Models Using Randomization. This process has been previously described as a method to assess whether the model generated is a random occurrence (Ekins et al., 2000a). Using the catScramble software in Catalyst with the training set, 10 validation trial sets were randomly produced in which activity was randomized with structure. These validation sets were then used as inputs for hypothesis generation. The resultant hypotheses generated with randomized data were assessed and the mean r² value calculated.

In Vitro Test Set Data and Pharmacophore Evaluation. The antipsychotic-derived hERG pharmacophore generated previously (Crumb et al., 2001) was evaluated with eight further molecules besides olanzapine, desmethyl-olanzapine, and 2-hydroxy methyl olanzapine (Table 1), which had been used previously (Crumb et al., 2001). All IC₅₀ values were derived using the cDNA-expressed hERG channels in HEK-293 cells at physiological temperature as previously described (Crumb, 2000). These test set molecules were fit by the fast algorithm method to the Catalyst hypothesis, to predict an IC₅₀ value. Fast fit refers to the method of finding the optimum fit of the substrate to the hypothesis among all the conformers of the molecule without performing an energy minimization on the conformers of the molecule (Catalyst tutorials, release 4.0; Molecular Simulations). The pharmacophore was further tested with 15 literature molecules (Table 2). Conversely, the General hERG pharmacophore generated with 15 literature molecules was tested using the 11 molecules in the antipsychotic-derived hERG model and the original 11-molecule test set.

Statistical Evaluation of Test Set Predictions. Observed and predicted inhibition data were graphed (data not shown) and fit using Microsoft Excel 97 to calculate an r² value. These data was also analyzed using the nonparametric Spearman's rho test available in JMP 4.0.2 (SAS Institute, Inc., Cary, NC). This test represents a correlation coefficient computed on the rank order of the data values and not the values themselves. This test also provides a statistical significance result expressed as the p value, where a value of <0.05 is meaningful.

	Results

Top Abstract Introduction Experimental Procedures Results Discussion References

Testing hERG Antipsychotic Pharmacophore Model Derived from in Vitro Data. The previously published hERG pharmacophore derived from data on hERG inhibition by antipsychotics demonstrated an r² of 0.77 for log-transformed observed and predicted data (Crumb et al., 2001). This pharmacophore was then tested with 11 molecules for which in vitro IC₅₀ values were generated using the same experimental method for hERG inhibition (Table 1). The rank order of observed and predicted data was assessed using the Spearman's rho coefficient, which was a statistically significant value of 0.71 (p < 0.014). A second test data set derived from 15 literature molecules with hERG inhibition values obtained under similar conditions to those in the present study was used to further evaluate this model (Table 2). In this case, the rank order of observed and predicted data generated a statistically significant Spearman's rho coefficient of 0.76 (p < 0.0011).

Generating and Testing a General hERG Pharmacophore. A General Catalyst pharmacophore was generated using 15 molecules with literature data (Table 2). The model using these molecules (Fig. 1) possessed four hydrophobes and one positive ionizable feature (three-dimensional coordinates and interfeature distances are shown in Table 3). The log-transformed observed versus predicted hERG IC₅₀ values resulted in an excellent correlation (r² value) of 0.90. The General hERG model was validated using the combined IC₅₀ values generated in the present study as well as a previous study (Crumb et al., 2001), which were fit to the pharmacophore model using the fast fit paradigm to generate predictions for inhibition of hERG (Table 1). The pharmacophore model also allows visualization of the fit of a molecule to the model. For example Fig. 2 shows the fit of the model to 9-hydroxyrisperidone, which illustrates an identical observed and predicted IC₅₀ (1.3 µM). In this case, the molecule fits four of five features. When the test set correlation coefficient of observed and predicted IC₅₀ data are compared, the 22-molecule test set generated an r² = 0.83. When the model is used to rank order the molecules ability to inhibit hERG for this same test set, a significant Spearman's rho coefficient of 0.77 (p = 0.0001) was obtained.

View larger version (175K): [in this window] [in a new window]   Fig. 1.   Catalyst General hERG pharmacophore showing hydrophobic (cyan) and positive ionizable features (red). Distances shown are in Å.

View larger version (164K): [in this window] [in a new window]   Fig. 2.   9-Hydroxyrisperidone fitted to the Catalyst General hERG pharmacophore derived from literature data showing hydrophobic (cyan) and positive ionizable features (red). Observed and predicted IC50 were 1.3 µM.

Validation of General hERG Pharmacophore by Randomization. To further test the validity of the pharmacophore one approach is to randomize the structures and activity to assess whether pharmacophores could still be built. If the correlation of the observed and predicted data for these scrambled models is significant it suggests the original model may be due to chance. Ten randomized structure-activity data sets for the General hERG literature data set were generated and used for hypothesis generation using Catalyst. The mean r² value for these models was 0.18 (range, 0-0.76), considerably lower than the actual pharmacophore described above in which the r² value was 0.90. This lower mean r² value after randomization is indicative that the actual hypothesis selected is significant and unlikely due to chance_.

	Discussion

Top Abstract Introduction Experimental Procedures Results Discussion References

To reduce the high attrition of candidates in the latter stages of drug development, the removal of candidate compounds in the preclinical stages that are likely to have poor pharmacokinetic and toxicity profiles will increase the efficiency of nomination and likely success of the remaining molecules. In recent years, combining the results of in vitro models with computational approaches has lead to the generation of predictive computational filters for absorption, distribution, metabolism, and excretion (Ekins et al., 2000b) and toxicity (Ekins and Rose, 2002). Hence, the discovery and optimization of new drug candidates are becoming increasingly reliant upon the combination of experimental and computational approaches used early in the overall process.

In recent years a number of drugs (including the prokinetic agent cisapride and the antihistamine terfenadine) have been removed from the market for reasons including their undesirable prolongation of the QT interval and incidences of life-threatening arrhythmias under certain clinical situations. This potentially serious adverse effect is believed to be mediated via a potent blockade of the hERG potassium channel and has increased the importance in understanding the structure-activity relationships for molecules to block this potassium channel. With clarification of the pharmacophore (or toxicophore) for this channel comes the possibility of reducing the likelihood of this side effect in drug candidates and possible toxicity in vivo. Because no crystal structure for hERG exists at present due to its membrane-bound nature, homology models based on the template bacterial KcsA channel and site-directed mutagenesis work (Mitcheson et al., 2000) have been used to infer important amino acid residues likely to be involved in the inhibitor-channel interaction. The data collected using site-directed mutagenesis work suggest that the amino acid residues located in the S6 transmembrane domain F656 and Y652 and to a lesser extent V625 and G648 are important for interaction. The work of Mitcheson et al. (2000) presents an important step forward in understanding the structural specificity for the hERG channel. However, the homology model itself may be of limited use in screening databases for molecules likely to interact with this domain because it is unlikely to provide rapid prediction of the ability to bind and inhibit this channel. Another computational approach, namely, 3D-QSAR, provides rapid quantitative predictions for ligand binding interaction and was used in the current study.

Using IC₅₀ data generated previously with cDNA expressed hERG channels in HEK-293 cells, a pharmacophore model was derived using 11 antipsychotic agents (Crumb et al., 2001) that explains the likely structural features in common between potent inhibitors. This model was initially tested with olanzapine and two olanzapine metabolites and was able to provide a useful ranking of the IC₅₀ values. Eight additional molecules were combined with these three molecules to form a test set of 11 molecules. A further 15 molecules with literature-derived hERG IC₅₀ data were used as a second test set for the antipsychotic-derived pharmacophore. In both cases these test sets were rank ordered in a statistically significant manner using the antipsychotic-derived pharmacophore because the Spearman's rho values were 0.71 for the 11 molecules and 0.76 for the 15 literature molecules. A Spearman's rho value of 1 would be optimal and 0 would be random rank ordering.

The literature set of 15 molecules was also used in an attempt to produce a more general model of hERG inhibition than that obtained with the data from the antipsychotic agents. The pharmacophore generated with the literature data set contains four hydrophobes and one positive ionizable feature. This is slightly different to the antipsychotic-derived pharmacophore in that the ring aromatic feature in the latter model is replaced with a positive ionizable feature and there are fewer hydrophobes in the former model. Agreement between the published homology model and the hydrophobic features in both hERG pharmacophores may appear to coincide with the F656 and Y652 residues in the homology model, which are involved in - stacking with aromatic residues in the inhibitors (Mitcheson et al., 2000). This is visualized by the fitting of 9-hydroxyrisperidone to the literature pharmacophore (Fig. 2). In this fit, the aromatic region of 9-hydroxyrisperidone is fitted to a hydrophobe and the molecule is bent over so that the rings systems at both ends are almost parallel; however, as with any model it is debatable as to whether this is representative of the situation in vivo.

The General hERG pharmacophore model based on literature data described in this study was also assessed using a test set of molecules excluded from the model, namely, using the data generated in our own laboratories. In the light of the work presented in this report, it would appear that the pharmacophore is able to generate predictions for the 22 molecules that correlate well with observed values (r² = 0.83) and produces a statistically significant rank ordering as indicated by the Spearman's rho coefficient of 0.77. It should also be noted that on the whole, this model predicted some of the IC₅₀ values higher than experimentally observed (Table 2) although in a few cases, such as terfenadine, this situation is reversed (Fig. 3). However, it is important to understand that the model is able to correctly rank order the inhibition of parent molecules and metabolites as indicated by the highly significant Spearman's rho coefficient. For example, the model can distinguish between thioridazine and its metabolite mesoridazine; clozapine and clozapine N-oxide and N-desmethylclozapine; risperidone and 9-hydroxyrisperidone as well as olanzapine and two of its metabolites. Furthermore, the General hERG model can distinguish potent inhibitors of hERG, such as thioridazine, cisapride, and sertindole from weaker inhibitors, such as nicotine, desmethyl olanzapine, and 2-hydroxymethyl olanzapine, thereby enabling us to rank order molecules, which may be valuable in early drug discovery.

View larger version (168K): [in this window] [in a new window]   Fig. 3.   Terfenadine fitted to the Catalyst General hERG pharmacophore derived from literature data showing hydrophobic (cyan) and positive ionizable features (red). Observed IC50 was 0.213 µM and predicted IC50 was 0.023 µM.

A major concern in understanding the clinical significance of hERG inhibition is the discrepancy observed between IC₅₀ values for hERG inhibition determined for the same molecules in different laboratories. For example, clozapine (Tie et al., 2000; this study), thioridazine (Tie et al., 2000; this study), and cisapride (Mohammad et al., 1997; Rampe et al., 1997; Crumb and Cavero, 1999; Walker et al., 1999a) show significant interlaboratory variability, which in many cases may be greater than 1 log unit. Further examples with such large IC₅₀ differences were also noted for ketoconazole (Dumaine et al., 1998; this study), haloperidol (Suessbrich et al., 1997; Crumb and Cavero, 1999; this study), and nicotine (Wang et al., 1999; this study). However, data for ketoconazole, haloperidol, and nicotine were generated in hERG expressed in both oocytes and HEK-293 cells, which may in part explain the interlaboratory difference in IC₅₀ values. Interestingly, sildenafil previously identified as a hERG inhibitor with an IC₅₀ of 100 µM in HEK-293 cells (Geelen et al., 2000) was shown in the current study with HEK-293 cells to have an IC₅₀ of 3.3 µM for hERG. This difference in IC₅₀ value derived with the same in vitro systems for hERG inhibition but between different laboratories suggests some influence due to different experimental procedures. Interestingly, the general hERG pharmacophore described in the present study indicated an IC₅₀ of 0.81 µM for sildenafil, whereas the antipsychotic-derived pharmacophore predicted an IC₅₀ of 0.18 µM (Table 2).

Protein binding and drug metabolism (neither of which are evaluated in this study) may also be important factors to consider in selecting molecules in addition to hERG inhibition. To some extent it has been suggested that hERG channel inhibition is not a class effect, at least in the case of fluoroquinolones (Kang et al., 2001). This recent study of seven antibiotics with IC₅₀ values for hERG inhibition in oocytes expressing this potassium channel produced a range of IC₅₀ values from 18 to 1420 µM. These in vitro data were also used in conjunction with free plasma concentrations to calculate hERG IC₅₀/plasma concentration ratios. The results indicate that some molecules known to prolong the QT interval in humans such as grepafloxacin possess low ratios, whereas other molecules such as ciprofloxacin have much higher ratios and consequently these have not been shown to prolong the QT interval.

By using and testing pharmacophore models built with our own data we have shown that literature data are as well predicted as our own when the rank order data are compared. At least in this study there appears to be little effect of potential interlaboratory differences that might impact model parameters and prevent model generation. Clearly, the General hERG pharmacophore we obtained differs slightly to that previously described for the antipsychotic pharmacophore because we have a ring aromatic or positive ionizable feature unique to both models. These pharmacophore features may represent important molecular interactions. Conversely, the fact that both models consist of multiple hydrophobes could be a consequence of the generally long flexible molecules in the present data sets. These observations suggest there may be multiple binding interactions within the potassium channel making this ultimately difficult to predict and could account for some of our poor predictions. Overcoming this limitation may require multiple pharmacophores or platform-specific models (like the antipsychotic-derived model described previously; Crumb et al., 2001) that could detect subtle structural differences and counteract the multiple conformations explored by Catalyst. To some extent literature studies have explored platform-specific models. The study on seven antibiotics with measured hERG IC₅₀ values (Kang et al., 2001) suggested the most potent hERG inhibitors contained C₅ substituents. It is unknown how this structure-activity relationship relates to other molecules from different therapeutic areas, which might limit its applicability. Some noncardiac drugs known to be hERG inhibitors have been suggested to contain the same structural feature pharmacophore as class III antiarrhythmics (a para-substituted phenyl ring connected to a basic nitrogen by a variable chain), whereas others do not (De Ponti et al., 2000). The hERG pharmacophores described to date may also have some additional value in the discovery and design of novel therapeutically useful drugs that are potent hERG inhibitors. Examples include class III antiarrhythmics that could prolong the action potential duration by potassium channel blockade (Rees and Curtis, 1996) and possible antiepileptics that might inhibit hERG expressed in the brain (Taglialatela et al., 1998).

In summary, in vitro IC₅₀ data obtained from the literature for the inhibition of the potassium channel encoded by the hERG gene in expressed cell systems can be readily used to a build a Catalyst computational model that possesses a predictive ability for many molecules excluded from the training set. Such a model may be subsequently used to rank molecules for their potential to inhibit this potassium channel and enable selection of molecules with relatively low ability to cause this interaction in vivo. As more data are generated and more sophisticated pharmacophores or computational models are developed beyond the preliminary ones described in our studies, we would also expect the quality of quantitative predictions to improve. Therefore, a preliminary in silico virtual screen to assess molecules for hERG inhibition has been developed that has been a goal suggested by numerous studies for preclinical development (De Ponti et al., 2000; Mitcheson et al., 2000).