WIXLIB

Author's personal copyA.H. Al-Nadaf, M.O. Taha / Journal of Molecular Graphics and Modelling 29 (2011) 843–864CATALYST requires informative training sets that include at least16 compounds of evenly spread bioactivities over at least threeand a half logarithmic cycles [43,45,47]. Four structurally diversetraining subsets (Table 3) were carefully selected from the collectedcompounds for pharmacophore modeling.Each training subset was utilized to conduct four modeling runsto explore the pharmacophoric space of renin inhibitors. Differenthypotheses were generated by altering the interfeature spacing andthe number of allowed features in the resulting pharmacophores(see Table B under Supplementary Materials).Eventually, our pharmacophore exploration efforts (16 automatic runs, Table 3 and Table B in Supplementary Materials)culminated in 160 pharmacophore models of variable qualities (seeSM-1 under Supplementary Materials for details about CATALYSTpharmacophore generation algorithm [43,45,47]).2.1.5. Assessment of the generated hypothesesWhen generating hypotheses, CATALYST attempts to minimizea cost function consisting of three terms: weight cost, error cost andconfiguration cost (see SM-2 pharmacophore assessment in CATALYST under Supplementary Materials [41,45,48]). In a successfulautomatic modeling run, CATALYST ranks the generated modelsaccording to their total costs [41].CATALYST also calculates the cost of the null hypothesis, whichpresumes that there is no relationship in the data and that experimental activities are normally distributed about their mean.Accordingly, the greater the difference from the cost of the nullhypothesis, the more likely that the hypothesis does not reflect achance correlation.CATALYST-HYPOGEN implements additional approach to assessthe quality of resulting pharmacophores, namely, the Cat-ScrambleOOOONH2NOH 2NHOO845R1NHNOO12 -3HNNH 4OR1OR2R135-50Fig. 2. The chemical scaffolds of training compounds, the corresponding structures and bioactivities are as in Table A in Supplementary Materials.

Author's personal copy846A.H. Al-Nadaf, M.O. Taha / Journal of Molecular Graphics and Modelling 29 (2011) 843–864H2NNR1NNNH2FOOOR4OO51-63HNOOR1OO64 -8 8HNO R1NOOOOOOOHNOR3R1NFNR2HNN89-9495 -9 8HNR1ONR1NOOOOOOO9 9 - 10 7108-11 1HNR1NHOOO112-119Fig. 2. (Continued).approach. This validation procedure is based on Fisher’s randomization test [47]. In this test, a 95% confidence level wasselected, which instructs CATALYST to scramble the bioactivitiesof training compounds to generate 19 random spreadsheets. Subsequently, CATALYST-HYPOGEN is challenged to use these randomspreadsheets to generate hypotheses using exactly the same features and parameters used in generating pharmacophore modelsfrom unscrambled bioactivity data. Success in generating pharmacophores of comparable cost criteria to those produced by theoriginal unscrambled data reduces the confidence in the training

Author's personal copyA.H. Al-Nadaf, M.O. Taha / Journal of Molecular Graphics and Modelling 29 (2011) 843–864compounds and their respective pharmacophore models [41,59].Only 62 pharmacophores, out of 160 generated models, werefound to possess Fisher confidence values 85%. Table C in Supplementary Materials shows the success criteria of representativepharmacophores from each run.2.1.6. Clustering of the generated pharmacophore hypothesesThe successful models (62) were clustered into 12 groupsutilizing the hierarchical average linkage method available in CATALYST. Therefore, closely related pharmacophores were groupedin five-membered clusters. Subsequently, the highest-ranking representatives, as judged based on their fit-to-bioactivity correlationr2 -values (calculated against collected compounds 1–119), wereselected to represent their corresponding clusters in subsequentQSAR modeling (Table C in Supplementary Materials).2.1.7. QSAR modelingA subset of 96 compounds from the total list of inhibitors(1–119) was utilized as a training set for QSAR modeling. However,since it is essential to assess the predictive power of the resulting QSAR models on an external set of inhibitors, the remaining23 molecules (ca. 20% of the dataset) were employed as an external test subset for validating the QSAR models. The selected testinhibitors are: 10, 11, 21, 33, 36, 45, 46, 48, 61, 65, 72, 76, 84, 88,101, 102, 106, 107, 108, 111, 114, 115 and 116 (numbers are as inTable A in Supplementary Materials and Fig. 2).The test molecules were selected as follows: the collectedinhibitors (1–119, Table A in Supplementary Materials and Fig. 2)were ranked according to their IC50 values, and then everyfifth compound was selected for the test set starting from thehigh-potency end. This selection considers the fact that the testmolecules must represent a range of biological activities similar tothat of the training set.The chemical structures of the inhibitors were imported intoCERIUS2 as standard 3D single conformer representations in SDformat. Subsequently, different descriptor groups were calculated for each compound employing the C2.DESCRIPTOR moduleof CERIUS2. The calculated descriptors were 128 properties (seesection SM-4 under Supplementary Materials) that included various simple and valence connectivity indices, electro-topologicalstate indices and other molecular descriptors (e.g., logarithm of partition coefficient, polarizability, dipole moment, molecular volume,molecular weight, molecular surface area, energies of the lowestand highest occupied molecular orbitals, etc.) [47]. Furthermore,the training compounds were fitted (using the Best-fit option inCATALYST) against the representative pharmacophores (12 models, Table C in Supplementary Materials), and their fit valueswere added as additional descriptors. The fit value for any compound is obtained automatically via equation (D) in SupplementaryMaterials [41].Genetic function approximation (GFA) was employed to searchfor the best possible QSAR regression equation capable of correlating the variations in biological activities of the training compoundswith variations in the generated descriptors, i.e., multiple linearregression modeling (MLR). The fitness function employed hereinis based on Friedman’s ‘lack-of-fit’ (LOF) [47]. However, to avoidoverwhelming GFA-MLR with large number of poor descriptors;we removed 50% of those showing lowest variance prior to QSARanalysis.We were obliged to normalize the potencies of the trainingcompounds via division by their corresponding molecular weights,i.e., ligand efficiency (Log(1/IC50 )/Mwt), to achieve reasonable selfconsistent QSAR models [35,49,50].Our preliminary diagnostic trials suggested the following optimal GFA parameters: explore linear, quadratic and spline equationsat mating and mutation probabilities of 50%; population size 500;847number of genetic iterations 30,000 and lack-of-fit (LOF) smoothness parameter 1.0. However, to determine the optimal numberof explanatory terms (QSAR descriptors), it was decided to scan andevaluate all possible QSAR models resulting from 4 to 20 explanatory terms.All QSAR models were validated employing leave one-out2 ), bootstrapping (r 2 ), leave 25%-out crosscross-validation (rLOOBS22validation (rL25%O ) and predictive r2 (rPRESS) calculated from therandomly selected external test subset (see selection criteria mentioned earlier).2In rL25%Oprocedure the training set is divided into two subsets:fit and test subsets. The test subset is randomly selected to represent 25% of the training compounds. This procedure is repeatedover four cycles; accordingly, four test subsets with their complementary fit subsets were selected for the particular QSAR model.The four test subsets should cover 100% of the training compoundsby avoiding selecting the same compound in more than one testsubset. The fit sets are then utilized to generate four QSAR submodels using the same descriptors. The resulting sub-models arethen utilized to predict the bioactivities of the corresponding testing subsets. Finally, the predicted values of all four test subsets arecorrelated with their experimental counterparts to determine the2corresponding rL25%O.2On the other hand, predictive rPRESSis defined as:2rPRESS SD PRESSSD(1)Where SD is the sum of the squared deviations betweenthe biological activities of the test set and the mean activity of the training set molecules, PRESS is the squareddeviations between predicted and actual activity valuesfor every molecule in the test set.2.1.8. Receiver operating characteristic (ROC) curve analysisSuccessful QSAR-selected pharmacophore models (i.e., Hypo1/5and Hypo1/7) were validated by assessing their abilities to selectively capture diverse renin inhibitors from a large list of decoysemploying ROC analysis.Therefore, it was necessary to prepare valid evaluation structural database (testing set) that contains an appropriate list of decoycompounds in combination with diverse list of known active compounds. The decoy list was prepared as described by Verdonk andco-workers [51,52]. Briefly, the decoy compounds were selectedbased on three basic one-dimensional (1D) properties that allowthe assessment of distance (D) between two molecules (e.g., i andj), namely: (1) the number of hydrogen-bond donors (NumHBD);(2) number of hydrogen-bond acceptors (NumHBA) and (3) countof nonpolar atoms (NP, defined as the summation of Cl, F, Br, I, Sand C atoms in a particular molecule). For each active compound inthe testing set, the distance to the nearest other active compoundis assessed using their Euclidean distance (equation (2)):D(i, j) 22(NumHBDi NumHBDj ) (NumHBAi NumHBAj ) (NPi NPj )2(2)The minimum distances are then averaged over all active compounds (Dmin). Subsequently, for each active compound in thetesting set an average of 20 decoys were randomly chosen fromthe ZINC database [53]. The decoys were selected in such a waythat they did not exceed Dmin distance from their correspondingactive compound.Moreover, to further diversify the actives members, i.e., to avoidclose similarity among actives in the testing set, any active compound having zero distance (D(i,j)) from other active compound(s)in the testing set were excluded. Active testing compounds weredefined as those possessing renin affinities (IC50 values) ranging

Author's personal copy848A.H. Al-Nadaf, M.O. Taha / Journal of Molecular Graphics and Modelling 29 (2011) 843–864from 0.067 nM to 100 nM. The testing set included 12 active compounds and 238 ZINC compounds.The testing set (250 compounds) was screened by each pharmacophore for ROC analysis employing the “Best flexible search”option implemented in CATALYST, while the conformational spacesof the compounds were generated employing the “Fast conformation generation option” implemented in CATALYST. Compoundsmissing one or more features were discarded from hit lists. Thein silico hits were scored employing their fit values (best-fit values) as calculated by equation (D) in Supplementary Materials.Subsequently, hit lists were used to construct ROC curves for corresponding pharmacophores (see section SM-3 ROC analysis inSupplmenetary Materials [51,54–56]).2.1.9. Addition of exclusion volumesTo account for the steric constraints of the binding pocket and tooptimize the ROC curves of our QSAR-selected pharmacophores, itwas decided to add exclusion volumes to Hypo1/5 and Hypo1/7employing the HIPHOP-REFINE module of CATALYST. HIPHOPREFINE uses inactive training compounds to add exclusion spheresto resemble the steric constraints of the binding pocket. It identifiesspaces occupied by the conformations of inactive compounds andfree from active ones. These regions are then filled with excludedvolumes [29–33,46].In HIPHOP-REFINE the user defines how many molecules mustmap the selected pharmacophore hypothesis completely or partially through controling the Principal and Maximum OmittedFeatures (MaxOmitFeat) parameters. Active compounds are normally assigned a MaxOmitFeat parameter of zero and Principalvalue of two to instruct the software to consider all their chemical moieties to fit them against all the pharmacophoric features ofthe particular hypothesis. On the other hand, inactive compoundsare allowed to miss one (or more) features by assigning them aMaxOmitFeat of one (or two) and Principal value of zero. Moderately active compounds are normally assigned a principal value ofone and a MaxOmitFeat of zero or one to encode their intermediatestatus.A subset of training compounds was carefully selected forHIPHOP-REFINE modeling. It was decided to consider IC50 of10 nM as an arbitrary activity/inactivity threshold. Accordingly,inhibitors of IC50 values 10 nM were regarded as “actives” andwere assigned principal and MaxOmitFeat values of two and zero,respectively, while less active inhibitors were assigned principalvalues of one or zero [29–33,46] and were carefully evaluatedto assess whether their lower potencies are attributable to missing one or more pharmacophoric features, i.e., compared toactive compounds (MaxOmitFeat 1 or 2), or related to possible steric clashes within the binding pocket (MaxOmitFeat 0).HIPHOP-REFINE was configured to allow a maximum of 150exclusion spheres to be added to the generated pharmacophorichypotheses.2.2. Fluorometric quantification of renin activitySensoLyteTM 520 Renin Assay Kit was used for the assay usinga Mc-Ala/Dnp FRET peptide [57]. The sequence of this peptideis derived from the cleavage site of renin. In this peptide, thefluorescence of Mc-Ala is quenched by Dnp. Upon cleavage intotwo separate fragments by renin, the fluorescence of Mc-Ala isrecovered, and can be monitored at excitation/emission s of490/520 nm.Test compounds and renin solutions were added into themicroplate wells and incubated at 37 C for 30 min. Subsequently,50 L renin substrate solution were added into each well. Thereagents were subsequently mixed thoroughly by shaking theplate gently for 30 s. The fluorescence intensity was immedi-ately measured continuously and recorded every 5 min for 15 minat 37 C. Appropriate positive and negative controls were prepared. The reaction rates were determined from the slopes ofabsorbance versus time plots constructed from 4 time points: 0, 510, and 15 min. Blank and standard inhibitor (Ac-His-Pro-Phe-ValSta-Leu-Phe-NH2) [57] were used as negative and positive controls,respectively.3. Results and discussionCATALYST enables automatic pharmacophore construction byusing a collection of molecules with activities ranging overa number of orders of magnitude. CATALYST pharmacophores(hypotheses) explain the variability of bioactivity with respect tothe geometric localization of the chemical features present in themolecules used to build it. Different hypotheses were generatedfor a series of renin inhibitors. A total of 119 compounds wereused in this study (Fig. 2 and Table A in Supplementary Materials).Four training subsets were selected from the collection (Table 3).Each subset consisted of inhibitors of wide structural diversity.The biological activity in the training subsets spanned from 3.5to 4.0 orders of magnitude. Genetic algorithm and multiple linear regression statistical analysis were subsequently employedto select an optimal combination of complementary pharmacophores capable of explaining bioactivity variations among allinhibitors.3.1. Data mining and conformational coverageThe literature was surveyed to collect as many structurallydiverse renin inhibitors as possible. However, the fact that pharmacophore and QSAR modeling necessitates that the trainingcompounds should have been assayed by a single bioassay procedure restricted us to certain published inhibitors (1–119, seeTable A in Supplementary Materials and Fig. 2) [20–22,37–40].Nevertheless, in order to assess the structural diversity of the collected compounds, and hence their aptness for pharmacophoreand QSAR modeling, we calculated several diversity-related parameters for the collected list and compared them with closelyrelated list of compounds extracted from the ZINC database(329 compounds) [53], which we used as decoys for ROC analysis (see Section 2.1.8). Table 1 summarizes the calculatedparameters of the two lists together with the implemented calculation methodologies. The comparison in Table 1 suggests thatour training set is much more diverse than most QSAR training sets, which are normally very limited in their structuralmodifications.Furthermore, clustering through maximal dissimilarity partitioning using several physicochemical descriptors and connectivityfingerprints classified the collected compounds into 12 differentclusters (Table 2), which further points the chemical diversity of thecollected list. Molecular diversity is essential for efficient pharmacophore and QSAR modeling, and to unveil different binding modesassumed by diverse binding ligands.The 2D structures of the inhibitors were imported into CATALYST and converted automatically into plausible 3D singleconformer representations. The resulting 3D structures were usedas starting points for conformational analysis and in the determination of various molecular descriptors for QSAR modeling.The conformational space of each inhibitor was extensively sampled utilizing the poling algorithm employed within the CONFIRMmodule of CATALYST [58] and via the “Best” module to ensureextensive sampling of conformational space. Efficient conformational coverage guarantees minimum conformation-related noiseduring pharmacophore generation and validation stages [58].

Author's personal copyA.H. Al-Nadaf, M.O. Taha / Journal of Molecular Graphics and Modelling 29 (2011) 843–864Table 1Molecular diversity among training compounds (1–119) compared to diversitywithin the decoy molecules in the ROC list (see Section 2.1.8).Diversity parametersCollectedcompoundseNormalized number of assembliesaNormalized number of fingerprint featuresbMinimumFingerprint 00.980.87Property ecoys in theROC listfaDefined as the total number of Murcko assemblie

a Department of Medicinal Chemistry and Pharmacognosy, Faculty of Pharmacy, Applied Science University, Amman, Jordan b Drug Discovery Unit, Department of Pharmaceutical Sciences, Faculty of Pharmacy, University of Jordan, Amman, Jordan article info Article history: Received 11 July 2010 Received in revised form 31 January 2011 Accepted 3 .