3844
`
`J. Med. Chem. 2002, 45, 3844-3853
`
`Toward a Pharmacophore for Drugs Inducing the Long QT Syndrome: Insights
`from a CoMFA Study of HERG K+ Channel Blockers
`
`Andrea Cavalli,† Elisabetta Poluzzi,‡ Fabrizio De Ponti,‡ and Maurizio Recanatini*,†
`Department of Pharmaceutical Sciences, University of Bologna, Via Belmeloro 6, I-40126 Bologna, Italy, and
`Department of Pharmacology, University of Bologna, Via Irnerio 48, I-40126 Bologna, Italy
`
`Received March 25, 2002
`
`In this paper, we present a pharmacophore for QT-prolonging drugs, along with a 3D QSAR
`(CoMFA) study for a series of very structurally variegate HERG K+ channel blockers. The
`blockade of HERG K+ channels is one of the most important molecular mechanisms through
`which QT-prolonging drugs increase cardiac action potential duration. Since QT prolongation
`is one of the most undesirable side effects of drugs, we first tried to identify the minimum set
`of molecular features responsible for this action and then we attempted to develop a quantitative
`model correlating the 3D stereoelectronic characteristics of the molecules with their HERG
`blocking potency. Having considered an initial set of 31 QT-prolonging drugs for which the
`HERG K+ channel blocking activity was measured on mammalian transfected cells, we started
`the construction of a theoretical screening tool able to predict whether a new molecule can
`interact with the HERG channel and eventually induce the long QT syndrome. This in silico
`tool might be useful in the design of new drug candidates devoid of the physicochemical features
`likely to cause the above-mentioned side effect.
`
`Introduction
`The long QT syndrome (LQTS) is characterized by the
`prolongation of the QT interval of the surface electro-
`cardiogram and is associated with an increased risk of
`torsades de pointes, a ventricular tachyarrhythmia that
`may degenerate into ventricular fibrillation and sudden
`death.1 Several congenital and acquired disorders can
`lead to prolongation of the QT interval; of special
`interest is the fact that numerous agents, belonging to
`different drug classes, have been associated with QT
`prolongation and occurrence of torsades de pointes.
`Recently, several regulatory interventions have involved
`drugs for which this potentially fatal risk was recog-
`nized only after marketing authorization.2 Screening
`methods for the early detection of an effect on the QT
`interval are therefore required during the drug develop-
`ment process, and several in vivo and in vitro methods
`are now available.3,4
`The QT interval is defined as the time interval
`between the onset of the QRS complex and the end of
`the T wave and therefore includes both the ventricular
`depolarization and repolarization intervals. Although
`several pathophysiological mechanisms can lead to
`prolongation of the QT interval,5-7 the key mechanism
`for drug-induced QT prolongation is the increased
`repolarization duration through blockade of outward K+
`currents (especially the delayed rectifier repolarizing
`current, IK). In particular, most of the QT-prolonging
`drugs have been shown to inhibit the K+ channels
`encoded by the human ether-a`-go-go related gene
`(HERG), at the basis of the rapid component of IK named
`IKr.8-12 HERG K+ channel blockade is therefore the most
`
`* To whom correspondence should be addressed. Phone: +39 051
`2099720. Fax: +39 051 2099734. E-mail: mreca@alma.unibo.it.
`† Department of Pharmaceutical Sciences.
`‡ Department of Pharmacology.
`
`important mechanism through which QT-prolonging
`drugs increase cardiac action potential duration.13
`Notably, blockade of HERG K+ channels forms the basis
`of the therapeutic effect of class III antiarrhythmic
`drugs, but for all other drugs, it is an unwanted side
`effect that must be detected as early as possible during
`drug development. IC50 for inhibition of HERG K+
`channels expressed in different cell lines is considered
`a primary test to study the QT-prolonging potential of
`a compound.14
`Since none of the existing in vitro tests to assess the
`QT-prolonging potential of a compound has an absolute
`predictive value,4 the availability of in silico methods
`in the early phase of drug development would dramati-
`cally increase the screening rate and would also lower
`the costs compared to experimental assay methods. At
`the present time, to our knowledge, no theoretical
`screening method for the evaluation of QT-prolonging
`properties of molecules is publicly available. In an
`attempt to explore the possibility of building an in silico
`screening system based on the known structure-activ-
`ity relationships (SAR) of LQTS-inducing noncardiac
`drugs, we report here a pharmacophore for such drugs,
`based on an initial set of molecules taken from an
`organized list of QT-prolonging compounds.15 Further-
`more, we present a 3D QSAR model obtained by means
`of the CoMFA technique that attempts to correlate the
`physicochemical features of the drug molecules with
`their blocking activity toward the HERG K+ channel.
`The obtained CoMFA model was validated by predicting
`the biological activities of a set of molecules not used in
`deriving the 3D QSAR equations. We propose this model
`as a first step toward the development of a tool able to
`predict whether a molecule of pharmacological interest
`bears structural features likely to induce LQTS.
`
`10.1021/jm0208875 CCC: $22.00 © 2002 American Chemical Society
`Published on Web 07/25/2002
`
`1 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`HERG K+ Channel Blockers
`
`Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18 3845
`
`Chart 1. QT-Prolonging Drugs Considered in the Papera
`
`2 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`3846 Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18
`
`Cavalli et al.
`
`Chart 1 (Continued)
`
`3 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`HERG K+ Channel Blockers
`
`Chart 1 (Continued)
`
`Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18 3847
`
`a In the CoMFA analysis, compounds 1-31 and 32-37 were the training and the test sets, respectively.
`
`Methods
`1. QT Prolonging Drugs Selected for the Study.
`In principle, a pharmacophore accounting for a certain
`pharmacological action should be built by taking into
`consideration all (or most) of the compounds able to
`elicit such an action. In the case of the QT-prolonging
`effect, besides the class III antiarrhythmic drugs,
`several other agents are known to induce this response
`and new entries are reported almost monthly at an
`accelerated pace. Indeed, prolongation of the QT interval
`by nonantiarrhythmic drugs is not an unusual finding,
`but the yet-unanswered question is to what degree QT
`prolongation has to be considered clinically significant.
`In this regard, in recent papers, we proposed a list of
`noncardiac QT-prolonging drugs, based on specific clini-
`cal and nonclinical criteria, and aimed to make a
`starting point to maintain a “consensus list” to be
`periodically updated.15 It seemed reasonable to consider
`this database (about 140 compounds at publication time)
`as the source of molecules to be used in the construction
`of the pharmacophore. However, since the pharmacoph-
`oric scheme must refer to a well-defined molecular
`target (i.e., the HERG K+ channel), we selected from
`the whole list only those drugs for which HERG K+
`channel inhibition had been reported. The final set of
`molecules shown in Chart 1 was obtained by combining
`our list15 with Fenichel’s database,16 which also includes
`drugs used as antiarrhythmics; we selected only those
`drugs for which IC50 values for inhibition of HERG K+
`channels expressed in mammalian cells (HEK, CHO,
`COS, neuroblastoma cells; see Table 1) were available.
`We decided not to include IC50 values obtained in
`nonmammalian cell lines, such as Xenopus laevis oo-
`cytes, since it is now recognized that the use of these
`systems leads to a significant underestimation of a
`drug’s potency as a HERG K+ channel blocker (the
`highly lipophilic environment in Xenopus oocytes limits
`access of the drug to its site of action; experiments are
`carried out at 22 °C instead of 37 °C).14,17
`
`a
`
`Table 1. Observed and Calculated HERG K+ Channel
`Blocking Activity of Compounds 1-31
`pIC50obsd pIC50fit
`compound
`IC50 (nM)
`¢
`astemizole (1)
`0.9b
`0.51
`9.04
`8.53
`cisapride (2)
`0.23
`8.19
`7.96
`6.5b
`E-4031 (3)
`0.26
`8.11
`7.85
`7.7b
`9.5-15b
`dofetilide (4)
`0.24
`7.91
`7.67
`8.04 -0.19
`sertindole (5)
`7.85
`14b
`7.80 -0.06
`pimozide (6)
`7.74
`18b
`7.58 -0.03
`haloperidol (7)
`7.55
`28.1b
`7.82 -0.33
`droperidol (8)
`7.49
`32.2b
`thioridazine (9)
`7.45
`7.23
`0.22
`35.7b
`7.22 -0.33
`56-204b
`terfenadine (10)
`6.89
`7.05 -0.21
`verapamil (11)
`6.84
`143b
`6.88 -0.09
`domperidone (12)
`6.79
`162b
`loratadine (13)
`6.76
`5.83
`0.93
`173b
`6.81 -0.11
`halofantrine (14)
`6.70
`196.9c
`6.65 -0.20
`mizolastine (15)
`6.45
`350b
`6.30 -0.04
`bepridil (16)
`6.26
`550d
`azimilide (17)
`6.25
`6.15
`0.10
`560c
`mibefradil (18)
`5.84
`5.75
`0.09
`1430d
`chlorpromazine (19) 1470c
`5.83
`5.68
`0.15
`5.98 -0.51
`imipramine (20)
`5.47
`3400c
`5.64 -0.22
`granisetron (21)
`5.42
`3730b
`dolasetron (22)
`5.22
`4.99
`0.23
`5950b
`5.18 -0.08
`perhexiline (23)
`5.11
`7800b
`5.66 -0.66
`amitriptyline (24)
`5.00
`10000b
`5.02 -0.26
`diltiazem (25)
`4.76
`17300b
`18000-34400c
`sparfloxacin (26)
`4.58
`4.39
`0.19
`glibenclamide (27)
`4.13
`4.07
`0.06
`74000e
`4.35 -0.24
`50000-104000c
`grepafloxacin (28)
`4.11
`sildenafil (29)
`4.00
`3.50
`0.50
`100000b
`103000-129000c
`moxifloxacin (30)
`3.93
`3.82
`0.11
`4.16 -0.27
`gatifloxacin (31)
`3.89
`130000c
`a Calculated from the non-cross-validated CoMFA model. b In
`human embryonic kidney (HEK) cells. c In Chinese hamster ovary
`(CHO) cells. d In African green monkey kidney derived cell line
`COS-7. e In neuroblastoma cells.
`Compounds 1-31 were used for the pharmacophore
`generation, and they also constituted the training set
`for the CoMFA procedure. In this connection, for each
`drug, the HERG blocking potency expressed as an IC50
`value is reported in Table 1. It is remarkable that the
`drugs considered in this study span a potency interval
`as HERG K+ channel blockers of more than 5 log units.
`
`4 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`3848 Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18
`
`Cavalli et al.
`
`To assess the predictive ability of the CoMFA model,
`we considered a test set of molecules (32-37, Chart 1)
`taken from Fenichel’s list, or whose activity as a HERG
`K+ channel blocker was reported after the publication
`of our list (e.g., cocaine18,19).
`2. Molecular Modeling. Here, we describe the
`construction of the three-dimensional (3D) models of the
`molecules, the conformational search followed by a
`cluster analysis, the pharmacophore generation, and the
`3D QSAR analysis based on the comparative molecular
`field analysis (CoMFA) procedure.20 All the molecular
`modeling studies were carried out by means of the
`SYBYL21 and MacroModel22 software running on a
`Silicon Graphics workstation.
`2.1. Construction of the Models. The 3D models
`of the molecules were either retrieved23 from the
`Cambridge Structural Database (CSD),24 or modeled by
`adding functional groups on crystallographic skeletons.
`Thus, very few models were built by assembling frag-
`ments retrieved from the standard SYBYL library. The
`following molecules were directly retrieved from the
`CSD (the CSD code is in parentheses): astemizole, 1
`(ZENREP); cisapride, 2 (KEYOB); pimozide, 6 (PI-
`MOZD); haloperidol, 7 (HALOPB); droperidol, 8 (KA-
`MCIK); thioridazine, 9 (WAVCEB); verapamil, 11
`(CURHON); domperidone, 12 (BEQJUC); loratadine, 13
`(YOVZEO); halofantrine, 14 (SATRAG); bepridil, 16
`(CEZBAK); chlopromazine, 19 (DUKTOS); imipramine,
`20 (IMIPRB); perhexiline, 23 (DEPGUA01); amitript-
`yline, 24 (YOVZEO); diltiazem, 25 (CEYHUJ01); spar-
`floxacin, 26 (COQWOU); glibenclamide, 27 (DUNXAL);
`sildenafil, 29 (CAXZEG); risperidone, 34 (WASTEP);
`clozapine, 35 (CMPDAZ10); and cocaine, 36 (CO-
`CAIN10). The following molecules were built by adding
`fragments to crystallographic skeletons: E-4031, 3, was
`built starting from BERGUA; terfenadine, 10, from
`YIHJOO; mizolastine, 15, from CELNUG and LEJKUG;
`azimilide, 17, from LIAXBUD; mibefradil, 18, from
`BZDMAZ; granisetron, 21, from KUSZED; dolasetron,
`22, from KAMCIK; grepafloxacin, 28, and moxifloxacin,
`30, from NIVQAK; gatifloxacin, 31, from COQWOU;
`ziprasidone, 33, from DEDCIY, NUXWAE, and JIVAO.
`Norastemizole (32) and fexofenadine (37) are metabo-
`lites of astemizole (1) and terfenadine (10), respectively,
`and were built by modifying the structure of the parent
`compounds. Finally, only two molecules were built de
`novo, namely, dofetilide, 4, and sertindole, 5.
`The molecular models obtained were first energy-
`minimized by using steepest descent and conjugate
`gradient until a convergence of 0.005 kJ mol-1 Å-1 on
`the gradient was reached. Then, conformational searches
`were carried out in order to sample the potential energy
`surface. All of the classical molecular mechanics calcu-
`lations were performed by using the MMFF force
`field,25,26 which carries parameters for all the investi-
`gated molecules. Finally, the average conformers ob-
`tained from a cluster analysis (see below) were opti-
`mized by using the semiempirical Hamiltonian PM327
`as implemented in the SYBYL package (keywords
`GNORM ) 0.001, MMOK when needed).
`2.2. Conformational Search. The conformational
`space of each molecule was sampled by means of Monte
`Carlo analysis28 as implemented in the MacroModel
`software. In the Monte Carlo approach, the dynamical
`
`behavior of a molecule is simulated by randomly chang-
`ing dihedral angle rotations or atom positions. Then,
`the trial conformation is accepted if its energy has
`decreased from the previous one. If the energy is higher,
`various criteria can be applied to determine whether the
`new conformation should be accepted or not. In our
`simulations, all of the dihedral angles of single linear
`bonds were allowed to move freely, and the trial
`conformation was accepted if the energy was lower than
`that of the previous conformation or if the energy was
`within a fixed energy window (100 kJ/mol). The number
`of Monte Carlo trials was set equal to 7000. A high
`number, usually in the order of hundreds, of conformers
`was generated. To classify the conformations obtained
`for each molecule, a geometrical cluster analysis was
`carried out on the output of Monte Carlo searches.
`2.3. Cluster Analysis. Generally, a cluster analysis29
`provides for the most significant solutions, by using
`filtering screens based on one or more external criteria.
`Conformations were classified in terms of geometrical
`similarity, and in particular, two conformers were
`considered as belonging to the same family when the
`heavy atom’s root-mean-square displacement (rmsd)
`was lower than 1 Å. In this way, between 10 and 20
`families of conformers were generated for each molecule,
`and the average structure of each family was energy-
`minimized and used in the procedure of pharmacophore
`generation.
`2.4. Pharmacophore Generation. An inspection of
`molecules 1-31 of Chart 1 shows that they are widely
`varied from a structural point of view, which implies
`an objective difficulty in the search for a common
`pharmacophore. To tackle this problem, we adopted a
`“constructionist” approach to the pharmacophore gen-
`eration, consisting of the individuation of an initial
`reference structure (the template) onto which we over-
`lapped molecules starting from those with similar
`geometric and spatial characteristics. This superimposi-
`tion procedure led to the individuation of further phar-
`macophoric characteristics that were then used to add
`the most different molecules and to refine the initial
`alignment. The template chosen for this study was the
`crystal structure of astemizole (1), because this molecule
`is one of the most potent long QT-inducing drugs, and
`HERG channel blockers (IC50 ) 0.9 nM, Table 1). The
`astemizole crystal structure was directly retrieved from
`the CSD and then geometrically optimized by means of
`PM3. Initially, three pharmacophoric points on the
`astemizole molecule were defined, namely, the basic
`nitrogen of the piperidine cycle (N) and the centers of
`mass (centroids C0 and C1) of the two close aromatic
`moieties. However, considering that several molecules
`of the training set bear an aromatic group connected to
`the basic nitrogen, a fourth pharmacophoric point was
`defined as the centroid (C2) of the phenyl ring belonging
`to the N-(p-methoxyphenylethyl) substituent of astemi-
`zole. Onto this template, the best-fitting conformer
`(within 20 kJ/mol from the global minimum) of the other
`molecules was then superimposed by using the corre-
`sponding pharmacophoric functions.
`Not all the molecules displayed all four pharmacoph-
`oric points, and in such cases, the superimposition was
`based on the available points or it was further guided
`by other functions present on some molecules. Particu-
`
`5 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`HERG K+ Channel Blockers
`
`Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18 3849
`
`Table 2. Summary of the Pharmacophore Geometrical Parameters
`distances (Å)
`angles (deg)
`angles between planesa (deg)
`
`height above planesa (Å)
`
`C0-N-C2
`C0-N-C1
`C1-N-C2
`
`134-173
`30-65
`132-152
`
`P0-P1
`P0-P2
`P1-P2
`
`50
`160
`114
`
`N-P3
`C0-P2
`C1-P1
`C2-P0
`
`0.6
`1.7
`3.4
`1.2
`
`4.3-6.7
`C0-C1
`10.3-14.7
`C0-C2
`11.9-12.3
`C1-C2
`5.2-9.1
`C0-N
`5.7-7.3
`C1-N
`4.6-7.6
`C2-N
`a P0 is the plane containing C0, C1, and N. P1 is the plane containing C0, N, and C2. P2 is the plane containing C1, N, and C2. P3 is
`the plane containing C0, C1, and C2.
`larly, in the case of compounds 2, 5-8, 12-15, 17, and
`18, the halogen atom located in the para position on one
`phenyl ring (C0) was used to reinforce the fit. As regards
`the quinolones, which are the least potent HERG
`channel blockers considered in this study, their struc-
`tures were quite different from the structure of the
`template, which implied that they were superimposed
`to astemizole by first anchoring the molecular skeleton
`to the basic piperazine N atom. This oriented the
`centroid of the 4-piperidinone ring onto C0.
`In Figure 1a, the overall superimposition of com-
`pounds 1-31 is shown and the positions of the phar-
`macophoric points are indicated. In particular, the
`simplified pharmacophoric scheme shown in Figure 1b
`was obtained by connecting the average spatial positions
`of the pharmacophoric points of all the molecules.
`Distances and angles defining the pharmacophore ge-
`ometry are reported in Table 2.
`2.5. 3D QSAR through CoMFA. The most critical
`point in the CoMFA procedure is the alignment of the
`molecules in Cartesian space, which is usually ac-
`complished by following some hypothesis on the binding
`of compounds to the active site of their biological
`counterpart. In this study, the above-described phar-
`macophore based on astemizole was the starting point
`for the overall CoMFA alignment (training and test
`sets). Actually, the pharmacophoric superimposition was
`exploited to obtain the alignment for the 3D QSAR
`analysis, and in the cases where different conformations
`or orientations of the molecules were possible, the
`conformation/orientation giving the best statistical re-
`sults was chosen.
`A CoMFA table was built containing the biological
`activity data of the HERG channel blockers (the de-
`pendent variables, pIC50obsd; Table 1) and the values of
`steric and electrostatic fields at discrete points of the
`Cartesian space surrounding the molecules (the inde-
`pendent variables). The fields were generated by using
`an sp3 carbon atom with a formal charge of +1 as a
`probe. The region was generated automatically around
`the molecules by fixing a grid spacing of 1 Å. The
`statistical analysis was carried out by applying the PLS
`procedure to the appropriate independent variables and
`using the standard scaling method (COMFA_STD).
`Furthermore, to reduce the number of independent
`variables, an energy cutoff value of 30 kcal/mol was
`selected for both electrostatic and steric fields. The
`minimum (cid:243) for further filtering of the independent
`variables was set to 2.0 kcal/mol. Cross-validated PLS
`runs were carried out to establish the optimal number
`of components (the latent variables) to be used in the
`final fitting models. The number of cross-validated
`groups was always equal to the number of compounds
`(leave-one-out procedure), and the optimal number of
`
`Figure 1.
`(a) Orthogonal views of the superimposition of
`compounds 1-31 from which the pharmacophore was ob-
`tained. In the left view, the pharmacophoric frame is shown
`in magenta and the centroids of the pharmacophoric functions
`are indicated. (b) “QT pharmacophore” (geometric details are
`reported in Table 2).
`
`latent variables was chosen by considering the lowest
`standard error of prediction (scross).
`
`Results
`Pharmacophore. The pharmacophore for the QT-
`prolonging potential obtained from the drugs shown in
`Chart 1 (1-31) is depicted in Figure 1b. How this
`pharmacophore was derived from the set of molecules
`considered in the present work is described above and
`illustrated in Figure 1a. All conformers aligned repre-
`
`6 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`3850 Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18
`
`Cavalli et al.
`
`Table 3. Summary of the CoMFA Statistical Parameters
`0.767
`q2
`0.744
`scross
`172.87
`F
`0.952
`r2
`0.336
`s
`optimal number of components
`3
`steric field contribution
`0.716
`electrostatic field contribution
`0.284
`0.744
`r2
`a
`pred
`pred ) (SD - PRESS)/SD, where SD is the sum of the squared
`a r2
`deviations of each observed activity value for each molecule of the
`test set from the mean of the observed activity values of the
`training set (pIC50obsd values of Table 1) and PRESS is the sum of
`the squared deviations between predicted and observed values
`(¢ of Table 4).
`
`sent low-energy conformations of the molecules, and it
`can be seen that the final alignment shows a satisfactory
`superimposition of the pharmacophoric points. The
`pharmacophore is made by three aromatic moieties
`connected through a nitrogen function that is a tertiary
`amine throughout the whole set of molecules. This
`renders the compounds protonated at physiological pH,
`and the presence of the amine function in all the
`molecules might reasonably suggest the importance of
`a positively charged group in conferring biological
`activity. The three aromatic moieties of the pharma-
`cophore schematically represented by C0, C1, and C2
`are not carried by all the molecules. Actually, while C0
`is always present in the QT-prolonging drugs here
`considered, only some molecules bear the functions
`identified as C1 and C2. Moreover, a few drugs of the
`set also show a polar or polarizable function (mostly
`carboxylic or sulfonamidic groups, or halogen atoms,
`respectively) on the C0 aromatic ring (Figure 1a),
`seemingly responsible for further modulation of the
`activity.
`To better describe the pharmacophoric features, some
`geometric parameters are reported in Table 2. It is
`immediately apparent that the ranges of distances and
`angles between the points are rather wide, which might
`be a consequence of the relevant structural diversity of
`the set of molecules. However, the pharmacophoric
`scheme was drawn in the simplest way, in view of its
`application to a growing number of structurally diverse
`molecules.
`How all the chemical functions in the pharmacophore
`can affect the biological activity toward the HERG K+
`channel of QT-prolonging drugs will be discussed in
`light of the 3D QSAR analysis results.
`CoMFA Model. The results of the CoMFA analysis
`performed on the training set of the HERG K+ channel
`blockers 1-31 are shown in Table 3, where the main
`statistics for both cross-validated and non-cross-vali-
`dated PLS analyses are reported. The 3D QSAR model
`selected on the basis of the minimum scross value
`criterion has an optimal number of components equal
`to 3, and descriptive and predictive abilities are evalu-
`ated by the statistic parameters r2 ) 0.952 and s )
`0.336, and q2 ) 0.767 and scross ) 0.744, respectively.
`The predictive properties of the CoMFA model were
`more rigorously tested by calculating the HERG block-
`ing potency of a set of molecules not used in the
`derivation of the model (32-37, Chart 1). The result is
`pred ) 0.744, which is satisfactory. Note that despite
`an r2
`the small number of compounds considered in the test
`
`Figure 2. View of the steric and electrostatic CoMFA
`STDEV*COEFF contour maps. The regions where increasing
`the molecules’ volume increases HERG K+ channel blocking
`activity are green (0.028 level), and the region where increas-
`ing the volume decreases activity is yellow (-0.022 level). The
`electrostatic contours indicate an increase of activity with
`increasing positive (red, 0.010 level) and negative (blue, -0.012
`level) charge, respectively. The pharmacophoric frame is shown
`for reference.
`
`set, the variation of potency is rather high (about 3 log
`units), and it is recalculated by the model
`in an
`acceptable way, even if the extreme points (32 and 37)
`are not well predicted (but only 32 exceeds the scross
`value, Table 3).
`The CoMFA model accounting for the 3D QSAR of
`compounds 1-31 is illustrated by the contour maps
`shown in Figure 2. Sterically favorable regions (green)
`are located around the pharmacophoric points C1 and
`mostly C2, while the space around C0 seems sensitive
`to both steric and electrostatic properties of the mol-
`ecules. Particularly, the yellow contour indicates that
`increasing bulk is detrimental for the activity, while the
`red and blue contours (which are not, as it might appear,
`on the same line but show the blue volume pointing
`outward with respect to the red ones)
`indicate a
`prevalent favorable effect of positively and negatively
`charged groups, respectively.
`In Figure 3, the most (Figure 3a) and the least (Figure
`3b) potent HERG K+ channel blockers of the series
`(astemizole, 1, and gatifloxacin, 31, respectively) are
`shown along with the CoMFA contour maps in the
`orientation they assume in the CoMFA alignment. By
`inspection of the figures, it immediately appears that
`while relevant portions of astemizole indeed protrude
`within the favorable steric regions around C1 and C2
`(green), gatifloxacin does not make contact with such
`regions with any of its parts. On the contrary, this
`molecule partially reaches the forbidden steric contour
`
`7 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`HERG K+ Channel Blockers
`
`Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18 3851
`
`Figure 3. Different occupancy of the CoMFA regions by astemizole, 1 (a), and gatifloxacin, 31 (b).
`
`its carboxylate
`around C0 (yellow), and moreover,
`function falls close to the electrostatic red region over
`C0, where positively charged groups would be required
`for an optimal HERG K+ channel blocking activity.
`
`Discussion
`In this study, we present a pharmacophoric model for
`a series of structurally different molecules able to induce
`the LQTS. This pharmacophore (Figure 1b and Table
`2) covers a number of prototypical structures recognized
`to be responsible for the LQTS. The presence of four
`relevant pharmacophoric points allowed us to allude to
`more than classical pharmacophoric geometrical ele-
`ments (such as distances and angles between points).
`Actually, in our “QT pharmacophore”, angles between
`the three different planes generated by the four phar-
`macophoric points, and heights of the pharmacophoric
`points above the planes were also identified. In prin-
`ciple, one might suggest that these additional geo-
`metrical elements should be taken into consideration
`when designing new drugs, to avoid the synthesis of
`molecules bearing such geometrical features and to
`minimize the risk of induced LQTS. On the other hand,
`we want to emphasize the “in progress” character of the
`present pharmacophore because of the growing number
`of new molecules endowed with this undesired biological
`property. Consequently, as the diversity of the molecular
`structures associated with LQTS increases, new chemi-
`cal and geometrical features will have to be taken into
`account in the continuous process of pharmacophore
`building and refining. For instance, there is some not
`yet clearly defined crowding of polar groups in the
`
`region between N and C2 (Figure 2a), which might
`eventually result in the identification of a further
`pharmacophoric feature.
`In an effort to refine the pharmacophore description
`by providing at the same time a tentatively predictive
`tool, we then performed a CoMFA analysis on the same
`set used for the construction of the pharmacophore,
`employing the HERG K+ channel blocking data of the
`molecules as the dependent variable. This implied that
`we assumed the HERG K+ channel blockade as the
`molecular mechanism through which drugs 1-31 cause
`the prolongation of the QT interval. We are aware of
`the fact that HERG K+ channel blockade is the main,
`but not the exclusive, mechanism leading to QT interval
`prolongation;5-7 however, for the compounds taken into
`consideration, experimental data exist confirming such
`hypothesis.15 On the other side, relating the LQTS-
`inducing potential of the compounds of Chart 1 to the
`measured action against a defined biological target
`allowed us to attempt an interpretation of the 3D QSAR
`at the molecular level.
`Caution must be exerted when using biological data
`from different sources in QSAR studies: actually, the
`IC50 values listed in Table 1 were obtained from differ-
`ent cell lines expressing the HERG K+ channel and this
`might, in principle, raise a question about the feasibility
`of the 3D QSAR investigation. This is the reason we
`carefully selected, for the training set, a relatively small
`number of compounds (31 out of a list of 140 drugs in
`the database), including only those drugs for which IC50
`values for the inhibition of HERG K+ channels ex-
`pressed in mammalian cells were available. We believe
`
`8 of 10
`
`PENN EX. 2038
`CFAD V. UPENN
`IPR2015-01836
`
`

`
`3852 Journal of Medicinal Chemistry, 2002, Vol. 45, No. 18
`
`Cavalli et al.
`
`Table 4. Observed and Calculated HERG K+ Channel
`Blocking Activity of Compounds 32-37
`compound
`IC50 (nM)
`PIC50obsd pIC50fit
`¢
`norastemizole (32)
`0.83
`7.55
`6.72
`28b
`-0.10
`ziprasidone (33)
`152b
`6.82
`6.92
`-0.20
`risperidone (34)
`163b
`6.79
`6.99
`clozapine (35)
`191b
`0.54
`6.72
`6.18
`4.400-7.200b
`-0.17
`cocaine (36)
`5.24
`5.44
`-0.66
`fexofenadine (37)
`21570b
`4.67
`5.33
`a Calculated from the non-cross-validated CoMFA model. b HEK
`cells.
`
`a
`
`that this approach minimized the inhomogeneity of in
`vitro data and allowed us to use a larger number of
`compounds as a training set than would have been
`feasible using only our own experimental data for HERG
`K+ channel blockade. Indeed, the statistics of the model
`can be used as a guide to evaluate the significance of
`the results. As shown from the data of Table 3 and also
`from both the pIC50fit and ¢ values of Table 1, the QSAR
`model is reasonably good. Although it is not a definitive
`model, it can be used as the basis for further refining
`and continuous probing through the use of a larger
`number of compounds as more experimental data
`become available.
`Thus, from this study, a 3D QSAR model for the
`HERG K+ channel blocking potency was generated
`(Table 3). We acknowledge that lower standard devia-
`tion values (especially scross) would have been desirable.
`A reason for the rather high dispersion of the (cross-
`validated) predicted activity data might be related to
`the aforementioned inhomogeneity of the biological
`system where the experimental IC50 values were mea-
`sured. Different sensitivity of the cellular systems to the
`molecular perturbation caused by the drug can be
`responsible for some noise in the set of activity data not
`accounted for by the structural variation of the mol-
`ecules. Moreover, the possibility exists that the com-
`pounds under investigation bind HERG K+ channels in
`different states (i.e., open and close) and/or in different
`domains of the protein. As a consequence of that, the
`physicochemical requirements for the binding may
`change from one compound to another in a way not
`simply correlated to the structure of the molecule.
`Whether this might be the case also for loratadine
`(13, Table 1), which is severely underestimated by the
`model (¢ ) 0.93, i.e., more than twice the standard
`deviation of the non-cross-validated regression) is hard
`to say. Indeed, loratadine is one the most rigid molecules
`of the whole set 1-37, and this suggests an entropic
`factor (unaccounted for by the CoMFA field equations)
`possibly favoring its interaction with the channel pro-
`tein. However, the similar molecule amitriptyline (24,
`Table 1) is overestimated by the CoMFA model (¢ )
`-0.66), which might reflect an intrinsic difficulty of the
`model to treat rigid molecules (perhaps in the alignment
`step).
`Finally, the reliability of a QSAR model is usually
`assessed by testing its ability to predict the activity of
`an external set of molecul