Introduction

Flavoenzyme dihydroorotate dehydrogenase (DHODH) [EC 1.3.99.11] (Liu et al., 2004) is a fourth enzyme of pyrimidine de novo synthesis which catalyses oxidation of intermediate dihydroorotate (DHO) to orotate (ORO) (Fig. 1). Pyrimidines are required for the biosynthesis of DNA, RNA, glycoproteins and phospholipids (Jones, 1980). DHODH catalyzes synthesis of pyrimidines which are necessary for cell growth and proliferation of rapidly growing cells. Requirement of pyrimidine nucleotides are depended on cell type and developmental stage, involvement of de novo pathway is small in resting or fully differentiated cells where cells acquire pyrimidine mainly by the salvage pathways (Mascia et al., 2000). DHODH enzymes are divided into two families based upon their localization, amino acid sequence, substrate/cofactor dependence, and cellular localization (Norager et al., 2002; Bjornberg et al., 1999). Enzymes belongs to family-1 located in the cytosol, electron acceptors involved in second half reaction of redox process are either fumarate or NAD+ whereas family-2 enzymes transfer electrons to ubiquinone (CoQ), to which hDHODH belongs (Bjornberg et al., 1997). DHODH inhibitors blocks the growth of fast proliferating cell whereas cells which grow at normal speed can meet the requirement of pyrimidine bases from normal metabolic cycle. Inhibitors of hDHODH have proven efficacy for the treatment of cancer (Shawver et al., 1997; Baumann et al., 2009) and immunological disorders, such as rheumatoid arthritis and multiple sclerosis (Chen et al., 1986; Herrmann et al., 2004; Merrill et al., 2009). Brequinar (Shannon et al., 1999) and leflunomide (Fox et al., 1999; Rozman, 1998) (Fig. 2) are two examples of such compounds. Brequinar is an antitumor and immunosuppressive agent, while leflunomide, which is a prodrug of the active metabolite A771 726 (Williamson et al., 1996), shows immunosuppressive activity. We have recently compiled the literature pertaining recent advancements in the discovery and development of DHODH inhibitors (Vyas and Ghate, 2011). In search for selective hDHODH inhibitors, we have attempted 2D and 3D QSAR (CoMFA) studies on amino nicotinic acid and isonicotinic acid derivatives. QSAR models generated in this work can provide useful information for the design of new compounds with better hDHODH inhibitory activity.

Fig. 1
figure 1

Reactions catalyzed by DHODH

Full size image
Fig. 2
figure 2

DHODH inhibitors

Full size image

Material and method

2D QSAR was performed using Vlife MDS QSAR plus software and 3D QSAR (CoMFA) was performed using the SYBYL × 1.2 software from Tripos Inc., St. Louis, MO, USA on a HP computer with Core2 Duo processor and a window XP operating system.

2D QSAR modeling and data set

The hDHODH inhibitory activity data IC50 (μM) was taken from the published work of Castro Palomino Laria et al. (Castro et al., 2010). The negative logarithm of the measured IC50 (μM) values were converted to pIC50 and subsequently used as the dependent variable for QSAR study. Compounds were sketched using the 2D draw application and converted to 3D structures. Energy minimization and geometry optimization were conducted using Merck molecular force field (MMFF) and atomic charges, maximum number of cycles were 1000, convergence criteria (RMS gradient) was 0.01 and medium’s dielectric constant of 1 by batch energy minimization method. Conformational search was carried out by a systemic conformational search method. Energy minimized geometry was used for calculation of descriptors, a total of 208, 2D descriptors were calculated which encoded different aspects of molecular structure and consists of electronic, thermodynamic, spatial and structural descriptors, e.g., retention index (chi), atomic valence connectivity index (chiV), path count, chain path count, cluster, path cluster, element count, estate number, semi-empirical, molecular weight, molecular refractivity, logP, and topological index. Various alignment-independent (AI) descriptors were also calculated.

Selection of training and test set

Dataset of 26 molecules (Table 1) was divided into training (22) and test (4) set compounds. Selection of the training set and the test set molecules was done manually by considering the fact that test set molecules represent a range of biological activity similar to that of the training set. Thus, the test set was a true representative of the training set. This was achieved by arbitrarily setting aside four compounds as a test set with a regularly distributed biological data. Unicolumn statistics of test and training sets (Table 2) showed accurate selection of test and training sets, as maximum of the training set was more than that of test set and the minimum of training set was less than or equal to that of test set.

Table 1 Structure, experimental, and predicted activity with residual and 2D descriptors of amino nicotinic acid and isonicotinic acid derivatives
Full size table
Table 2 Unicolumn statistics of training and test sets (2D QSAR)
Full size table

Statistical computation

Vlife MDS was used to generate 2D QSAR models by multiple linear regression (MLR), principal component regression (PCR) and partial least squares (PLS) regression methods, coupled with forward–backward variable selection method. Statistical measures were used for the evaluation of 2D QSAR models were the number of compounds in regression n, regression coefficient r 2, number of descriptors in a model k, F-test (Fisher test value) for statistical significance F, cross validated correlation coefficient q 2, predictive squared correlation coefficients r 2 pred, coefficient of correlation of predicted data set pred_r 2se and standard error of estimation r 2 se and q 2 se.

Multiple linear regression (MLR) analysis

MLR is a regression method used to model linear relationship between a dependent variable Y (hDHODH inhibitory activity) and independent variables X (2D descriptors). MLR is based on least squares: the model is fit such that sum-of-squares of differences of observed and a predicted value is minimized. MLR estimates values of regression coefficients (r 2) by applying least squares curve fitting method. The model creates a relationship in the form of a straight line (linear) that best approximates all the individual data points. Regression equation takes the form

$$ Y \, = \, b1* x1 \, + \, b2* x2 \, + \, b3* x3 \, + \, c $$

where Y is dependent variable, ‘b’s are regression coefficients for corresponding ‘x’s (independent variable), ‘c’ is a regression constant, or intercept (Kubyani 1994; Croux and Joossens 2005).

Principal component regression (PCR) method

PCR is a data compression method based upon the correlation among dependent and independent variables. PCR provides a method for finding structure in data sets. Its aim is to group correlated variables, replacing the original descriptors by new set called principal components. These principal components uncorrelated and are built as a simple linear combination of original variables. It rotates the data into a new set of axes such that first few axes reflect most of the variations within the data. PCA selects a new set of axes for the data. These are selected in decreasing order of variance within the data. Purpose of PCR is the estimation of values of a dependent variable on the basis of selected principal components (PCs) of independent variables (Huberty, 1984).

Partial least squares (PLS) regression method

PLS analysis is a popular regression technique which can be used to correlate one or more dependent variable (Y) to several independent (X) variables. PLS relates a matrix Y of dependent variables to a matrix X of molecular structure descriptors. PLS is useful in situations where the number of independent variables exceeds the number of observation, when X data contain colinearties or when N is less than 5 M, whares N is number of compound and M is number of dependant variable. Main aim of PLS regression is to predict the activity (Y) from X and to describe their common structure (Wold et al., 2001).

2D QSAR models were generated using pIC50 values as dependent variable and various descriptors values as independent variables. The cross-correlation limit was set at 0.5, number of variables in final equation is 4 in MLR, PCR and PLS. Term selection criteria was set as r 2 and F-test, ‘in’ at 4 and ‘out’ at 3.99. Variance cutoff was set at 0, scaling to auto scaling and number of random iterations to 10.

Validation of QSAR models

The definitive validity of the model is examined by mean of external validation (q 2), which evaluates how well the equation generalizes. The training set was used to derive an adjustment model that was used to predict the activity of the test set. The predicted power of equations was validated using predictive squared correlation coefficients r 2 pred.

3D QSAR modeling and data set

The structures of all the compounds were constructed from the template molecule (compound 25) by using the “SKETCH” option function in SYBYL, and partial atomic charges were calculated by the Gasteiger Huckel method and energy minimizations were performed using the Tripos force field (Gasteiger and Marsili, 1980) with a distance-dependent dielectric and the Powell conjugate gradient algorithm convergence criterion of 0.01 kcal/mol Å (Clark et al., 1989). The total set of inhibitors was divided manually into training set of 22 compounds for generating 3D QSAR model and a test set of 4 compounds for validating the quality of the model.

Molecular modeling and alignment

The alignment of molecules is the process of aligning two or more molecules in 3D space to optimally superimpose specific atoms on each other based on distances. Compound 25 was used as a template because of the highest activity and all other compounds were aligned on the basis of the common structure (Fig. 3). Rigid body alignment of molecules in a Mol2 database was performed using maximum common substructures defined by Distill (without including bond types in rings). Structure of the template compound 25 and common substructure in bold is shown in Fig. 3. Alignment of training and test set compounds is shown in Fig. 4.

Fig. 3
figure 3

Structure of the template compound 25, common substructure is in bold

Full size image
Fig. 4
figure 4

Alignment of training and test set compounds on compound 25

Full size image

CoMFA model

CoMFA steric and electrostatic interaction fields of each molecule were calculated on a 3D cubic lattice with grid spacing of 2 Å in all the Cartesian directions and CoMFA fields were calculated using the QSAR module of SYBYL. CoMFA descriptors were calculated using sp3 carbon probe atom with a van der Waals radius of 1.52 Å and a charge of +1.0 to generate steric (Lennarde-Jones 6–12 potential) field energies and electrostatic (Coulombic potential) fields with a distance-dependent dielectric at each lattice point. The SYBYL default energy cutoff of 30 kcal/mol was set for both steric and electrostatic fields. In order to reduce noise and improve efficiency, column filtering (minimum sigma) was set to 2.0 kcal/mol.

Predictive r 2 value

To validate the CoMFA model, predictive abilities for the test set compounds (expressed as r 2 pred) was determined using the following equation

$$ {r^{ 2}}_{\text{pred}} = {\text{ SD}}-{\text{PRESS}}/{\text{SD}} $$

where, SD is the sum of the squared deviations between the inhibitory activity of molecules in the test set and the mean inhibitory activity of the training set molecules, and PRESS is the sum of the squared deviations between predicted and actual activity values for every molecule in the test set.

Analysis of the residuals

The training set was initially checked for outliers for both 2D and 3D QSAR analysis. In general, if the residual of a compound between experimental and predicted pIC50 values is greater than 1 logarithm unit, compound is considered as outlier. Examination of the residuals from cross-validated predictions (Tables 1, 3) indicated that there is no outlier in 2D and 3D QSAR models.

Table 3 Experimental and predicted pIC50 with residual values using 3D QSAR (CoMFA) model
Full size table

Results and discussion

Results of 2D QSAR study

Generation of 2D QSAR models

2D QSAR study on amino nicotinic acid and isonicotinic acid derivatives resulted in several QSAR models. Statistically significant QSAR models were selected for discussion.

Model-1 (MLR)

\( {\text{pIC}}_{ 50} = \, + \, 0. 4 500 \, \left( {{\text{T}}\_{\text{N}}\_{\text{F}}\_ 5} \right) \, + \, 0.0 6 1 3 { }\left( { 4 {\text{pathClusterCount}}} \right) \, - \, 0. 1 1 5 7 \,\left( {{\text{T}}\_{\text{C}}\_{\text{C}}\_ 6} \right) \, + { 6}. 7 6 4 5 \) where, n = 22training and 4test, k = 2, DF = 19, r 2 = 0.834, q 2 = 0.756, F test = 47.83, r 2 se = 0.243, q 2 se = 0.295, r 2 pred = 0.793, pred_r 2se = 0.334.

Model-2 (PCR)

\( {\text{pIC}}_{ 50} = \, + \, 0. 4 2 7 5 \, \left( {{\text{T}}\_{\text{N}}\_{\text{F}}\_ 5} \right) \, + \, 0.0 6 4 7 \left( { 4 {\text{pathClusterCount}}} \right) \, - \, 0. 1 1 5 6 \, \left( {{\text{T}}\_{\text{C}}\_{\text{C}}\_ 6} \right) \, + { 6}. 6 7 8 9 \) where, n = 22training and 4test, k = 2, DF = 19, r 2 = 0.833, q 2 = 0.774, F test = 47.34, r 2 se = 0.244, q 2 se = 0.284, r 2 pred = 0.811, pred_r2se = 0.304.

Model-3 (PLS)

\( {\text{pIC}}_{ 50} = \, + \, 0. 4 500 \,\left( {{\text{T}}\_{\text{N}}\_{\text{F}}\_ 5} \right) \, + \, 0.0 6 1 3 { }\left( { 4 {\text{pathClusterCount}}} \right) \, - \, 0. 1 1 5 7 \,\left( {{\text{T}}\_{\text{C}}\_{\text{C}}\_ 6} \right) \, + { 6}. 7 6 4 5 \) where, n = 22training and 4test, k = 2, DF = 19, r 2 = 0.864, q 2 = 0.786, F test = 48.83, r 2 se = 0.233, q 2 se = 0.294, r 2 pred = 0.821, pred_r 2se = 0.304.

In above QSAR models, r 2 is a correlation coefficient that multiply by one hundred gives explained variance in inhibitory activity. Predictive ability of generated QSAR models was evaluated by q 2 employing leave-one-out method. F value reflects ratio of variance explained by models and variance due to error in regression. High F value indicates that model is statistically significant. Low standard error (SE) of estimation indicted by r 2 se and q 2 se, suggested that models are statistically significant. Predictive ability of QSAR model was also confirmed by external validation of test set compounds denoted by r 2 pred. Observed and predicted pIC50 is shown in Table 1. Plot of observed versus predicted pIC50 is shown in Fig. 5.

Fig. 5
figure 5

Graphs of experimental versus predicted pIC50 using 2D QSAR models

Full size image

Interpretaion of 2D QSAR models

Descriptors used in generation of 2D QSAR models are shown in Fig. 6. 2D QSAR models indicates positive contribution of T_N_F_5 and 4pathClusterCount and negative contribution of T_C_C_6. Alignment independent (AI) topological descriptor (Balaban, 1982) T_N_F_5 contributed positively to QSAR models, where T_C_C_6 contributed negatively. Alignment-independent descriptors can be generated by considering the topology of the molecule, atom type, and bond. For calculation of alignment independent descriptors every atom in the molecule was assigned at least one and at most three attributes. First attribute is ‘T-attribute’ to thoroughly characterize topology of the molecule. Second attribute is atom type, atom symbol is used here. Third attribute is assigned to atoms taking part in a double or triple bond. After all the atoms have been assigned their respective attributes, selective distance count statistics for all combinations of different attributes are computed. A selective distance count statistic ‘XY2’ (e.g., ‘TOPO2N3’) counts all the fragments between start atom with attribute ‘X’ (e.g., ‘2’ double bonded atom) and end atom with attribute ‘Y’ (e.g., ‘N’) separated by graph distance 3. Graph distance can be defined as the smallest number of atoms along the path connecting two atoms in the molecular structure. In this study, to calculate AI descriptors, we used following attributes: 2 (double bonded atom), 3 (triple bonded atom), C, N, O, H, F, and Cl the distance range of 0–7. T_N_F_5 is a count of number of nitrogen atoms separated from any fluorine atom (single or double bonded) by five bond distance, e.g., N_C_N_C_C_C_F. Positive contribution of T_N_F_5 reveals the importance of presence of nitrogen atom in pyridine ring and fluorine atom on first phenyl ring of biphenyl ring template. T_C_C_6 is a count of number of carbon atoms separated from any other carbon atom (single or double bonded) by six distance, e.g., C_C_C_C_C_C_C_C. 4pathClusterCount is a molecular connectivity index which signifies total number of fragments of fourth order path cluster in a molecule. Molecular connectivity index is used to describe electronic environment and bonding configuration of each non-hydrogen atom (heavy atom) in the molecule for example carbon valence connectivity index takes into account only bonds between carbon atoms. 4pathClusterCount reveals the importance of molecular connectivity for heavy atoms and their bonding configuration in the molecules.

Fig. 6
figure 6

Contribution charts of 2D QSAR models

Full size image

Results of 3D QSAR study

The q 2, r 2 pred, r 2ncv , F, and SEE values were computed as defined in SYBYL. PLS analysis showed a high q 2 value of 0.630 with four components. Cross-validated q 2 (r 2cv ) of 0.630 indicated a good predictive ability of the model. The non-cross-validated PLS analysis results in a conventional r 2 of 0.949, F = 137 and a standard error of estimation (SEE) of 0.218. In both steric and electrostatic field contributions, the former accounts for 0.554, while the latter contributes 0.446, indicating that these two factors nearly contribute the same to the binding affinities. The high bootstrapped r 2 (0.966) value and low standard deviation (0.032) suggest a high degree of confidence in the analysis. The predicted, experimental activity and the residual value of all the inhibitors are listed in Table 3, and the correlation between predicted and experimental activity is depicted in Fig. 7. The predictive ability of the 3D QSAR model was further validated using an external test set of four compounds not included in the model generation study. The predicted r 2 (r 2 pred) values from the CoMFA model was 0.763.

Fig. 7
figure 7

Plot of experimental versus predicted pIC50 using 3D QSAR (CoMFA) model

Full size image

CoMFA contour maps

Contour maps for the best CoMFA model are shown in Fig. 8. In the contour maps, the steric CoMFA contour plot with the highest active compound 25 is shown in Fig. 8a. The field energies at each lattice point were calculated as the scalar results of the coefficient and the standard deviation associated with a particular column of the data table (std*coeff), as always plotted as the percentages of the contribution of CoMFA equation. In this figure, the green contours represent regions of high steric tolerance (80% contribution), while the yellow contours represent regions of low steric bulk tolerance (20% contribution). The steric contour of CoMFA showed a large green contour around the first phenyl ring of biphenyl ring template, indicating a favorable effect of steric bulk of fluorine atom for inhibitory activity. This steric favored area is generated by high electron density of fluorine atom. This can be explained by analyzing the structural features and inhibitory activity of 25 (2,3,5,6-tetrafluorophenyl, IC50 = 3 μm) and 26 (2,6-difluorophenyl IC50 = 11 μm), 13 (2-flouro, 5-methylphenyl IC50 = 99 μm), and 3(2-chlorophenyl IC50 = 150 μm). Fluorine atom is larger than hydrogen, thus steric bulk (lipophilicity) in the molecule can be increased by replacing H atom by F atom. A steric unfavorable yellow contour was observed near the C-3′ methoxy at terminal phenyl ring of biphenyl ring template, suggested that bulky groups in these region would decrease hDHODH inhibitory activity. CoMFA electrostatic contour map is shown in Fig. 8b. Regions where increased positive-charge is favorable for inhibitory activity are indicated in blue (80% contribution), while regions where increased negative-charge is favorable for inhibitory activity are indicated in red (20% contribution). A large region of red contours near the first phenyl ring of biphenyl ring template shows that the presence of electronegative substituent (–F). Fluorine has the highest electron density, thus such an electronegative groups (–CF3, –OCF3) are very important for better hDHODH inhibitory activity. It can shows the fact that activity of 13 (2-fluoro, 5-methylphenyl IC50 = 99 μm) is less than 2 (2,5-difluorophenyl IC50 = 88 μm). A large blue contour is seen in the vicinity of terminal phenyl ring, depicts that positively charged groups, such as hydrogen atoms is beneficial for inhibitory activity. This is indeed the case for 24 (IC50 = 8 μm) and 4 (IC50 = 90 μm). Second blue polyhedron near the C-2 position of pyridine ring indicate that a low electron density in this area will have a positive effect on the inhibitory activity. Small blue polyhedra located near the nitrogen atom of pyridine ring, indicate that an electropositive group needs to be present in this region.

Fig. 8
figure 8

CoMFA (std*coeff) contour maps. Compound 25 is shown inside the field, a CoMFA steric contour map and b CoMFA electrostatic contour maps

Full size image

2D versus 3D QSAR (CoMFA) analysis

The comparison of 2D and 3D QSAR (CoMFA) analysis suggested common structural features responsible for hDHODH inhibitory activity. Positive contribution of alignment-independent topological descriptor T_N_F_5 reveals the importance of nitrogen atom in pyridine ring and fluorine atom on first phenyl ring of biphenyl ring template separated by five bond distance. T_N_F_5 is an important descriptor, accounts for highest contribution (Fig. 6) for hDHODH inhibitory activity in all the 2D QSAR models. Fluorine is much more lipophilic than hydrogen, so incorporating fluorine atoms in a molecule will make it more lipophilic. Lipophilicity is an important property in describing the affinity of the compounds in terms of their partitioning the biological membranes hence the fluorinated compound has a higher bioavailability. Fluorine is a good leaving group, so it has a potential for covalent bonds to be formed between the molecule and hDHODH by loss of fluoride, leading to inhibition of hDHODH activity. Lone pair of electron on N-atom in pyridine ring system can form H-bond with CoQ binding site of hDHODH. Positive contribution of 4pathClusterCount reveals the importance of molecular connectivity for heavy atoms and their bonding configuration in the molecules. Analysis of CoMFA steric and electrostatic contour plots offered enough information to understand the binding mode between the inhibitors and binding site (CoQ) of hDHODH. The bulky and electronegative group (–F) of compound at the first phenyl ring of biphenyl ring system seems to be penetrating the junction of red (electrostatic) and green (steric) contours indicating the presence of bulkiness as well as electronegativity for the enhancement of hDHODH inhibitory activity. 2D and 3D QSAR models suggested that substitution on first phenyl ring especially with –F, –CF3, and –OCF3 and terminal phenyl ring with positively charged groups led toward better inhibitory activity.

Conclusion

2D and 3D QSAR study identifies common features responsible for hDHODH inhibitory activity of nicotinic acid and isonicotinic derivatives. 2D QSAR studies revealed that alignment-independent descriptors were major contributing descriptors. CoMFA model is satisfactory according to the statistical results as well as the contour maps analysis. CoMFA contour plots offered enough information to understand the binding mode between inhibitors and CoQ binding site of hDHODH. The most significant feature for better hDHODH inhibitory activity is the substitution pattern (–F, –CF3, –OCF3) on biphenyl ring system. QSAR models generated in this study can provide useful information for the design of new compounds and helped in prediction of hDHODH inhibitory activity prior to synthesis.