English

• 1. INTRODUCTION

  Neuraminidase (NA) inhibitors are the only class of antiviral drugs currently approved for use against the influenza virus^[1]. Influenza often occurs in winter seasons and occasionally causes pandemics due to the emergence of virus mutants and unexpected transmission to humans. The influenza virus, avian influenza A (H7N9) virus in particular, is a great threat to human health potentially causing serious public health and economic problems^[2]. Due to the key role in replication, spread, infection and pathogenesis of influenza virus, the application of neuraminidase inhibitors (NAIs) has been considered as a dominant approach for the treatment of influenza infection^[3], because the active site of NA is highly conserved across the influenza A and B viral strains^[4]. This inspired us to explore novel pyrimidine derivatives as NA inhibitors. Moreover, QSAR^[5] is widely used in drug design.

  The availability of computational techniques on quantitative structure activity relationship (QSAR) might provide a potential direction for accelerating the drug design process. In fact, QSAR can be viewed as a technique attempting to summarize chemical and biological information in a form that allows one to generate relationships between chemical structure and biological activity^[6]. As is well known, the success of a QSAR study depends also on the selection of variables (molecular descriptors) and on the representation of the information. Variables should give the maximum of information in active variables. 3D-QSAR model would better reflect the interaction between the ligand and receptor compared to 2D-QSAR. Three-dimensional holographic atomic vector field (3D-HoVAIF) is a widely used method for 3D-QSAR. In this paper, QSAR models of 28 thiazolidine-4-carboxylic acid derivatives as novel influenza NAIs were constructed to understand the chemical-biological interactions governing their activities toward influenza NA^[7]. This article by means of 3D-HoVAIF and MLR studied the 3D-QSAR models of molecular drugs, and then using molecular docking studied the combination of the active site between drugs and neuraminidase.

  2. PRINCIPLE AND METHODOLOGY

  In 3D-QSAR, 3D molecular fields were used as molecular descriptors. Assuming that the compounds considered in some applications bind to the target in marginally same way, the value of a 3D molecular field at every point in space was used as a descriptor. If then one assumes that the variation in biological activity can be related to the change in these 3D molecular field values between the molecules, a 3D-QSAR model can be obtained.

  3D-QSAR models are relatively simply interpretable through visualization where the important regions that correspond to descriptors retained in the QSAR are located. Once it is known in what regions these resides and what effect they have on QSAR, the ligand candidates to maximally exploit these regions and hence increase the (predicted) activity of the compounds were manipulated by drug designers. The first available 3D-QSAR analysis was carried out by 3D-HoVAIF which was proposed as a three-dimensional molecular structure description and multiple drug system has been applied^[8-11].

  3D-HoVAIF was generated considering three common non-bonding interactions of the biological activities, i.e., electrostatic, steric, and hydrophobic interaction related with atomic relative distance and atomic self-properties. These descriptors neither resort to any experimental parameters nor consider configuration overlap of samples. Ordinary atoms of organic molecules including H, C, N, P, O, S, F, Cl, Br and I are classified into five types in the Periodic Table of Elements. According to the hybridization state of atoms, the atoms are furthermore subdivided into ten types. Thus, there are 55 interactions in a molecule^[12]. In this paper, electrostatic, steric and hydrophobic potential energies take part in the representation of different interactions, producing 3 × 55 = 165 interaction items for organic molecules.

  Its electrostatic interactions can be used in the classical Coulomb theorem (type (1)) to describe;

  $ E_{mn} = \sum\limits_{i \in m, j \in n} {\frac{{{e^2}}}{{4\pi \varepsilon _0}}} \bullet \frac{{Z_i \bullet Z_j}}{{r_{ij}}} \;\;\;\;\;\;(1 \leqslant m \leqslant 10, m \leqslant n \leqslant 10) $

  (1)

  Where $ {r_{ij}} $ is the interatomic Euclid distance (nm), e the unit electric charge of 1.6021892 × 10^–19C, ε₀ the vacuum dielectric constant being 8.85418782 × 10^–12 C²/J·m, Z the amount of net electric charges; m and n are atomic types. The entire electrostatic interaction items are calculated by this formula. Fifty-five electrostatic interaction items are calculated by this formula according to their attributions.

  Stereoscopic effect of the Lennard-Jones equation to describe the effect of the type (type (2));

  $ S_{mn} = \sum\limits_{i \in m, j \in n} {{\varepsilon _{ij}}} \bullet D \bullet \left[ {{{\left( {\frac{{R_{ij}^ * }}{{{r_{ij}}}}} \right)}^{12}} - 2 \bullet {{\left( {\frac{{R_{ij}^ * }}{{{r_{ij}}}}} \right)}^6}} \right] \;\;\;\;(1 \leqslant m \leqslant 10, m \leqslant n \leqslant 10) $

  (2)

  Where ε_ij represents the potential well^{[13, 14]}, D is 0.01^[14], representing the calibration constant of interatomic interaction deduced by experience; R_ij^* = (C_h·R_ii^* + C_h·R_jj^*)/2 is Van derwaals radius with its calibration factor of 1.00 in the case of sp³, 0.95 in sp² and 0.90 in sp hybridization^[15].

  Hydrophobic action uses Hint of Kellogg and method to define the hydrophobic effect between the two atoms (type (3));

  $ H_{mn} = \sum\limits_{i \in m, j \in n} {{S_i}} \bullet {a_i} \bullet {S_j} \bullet {a_j} \bullet {e^{ - {r_{ij}}}} \bullet {T_{ij}} \;\;\;\;\;\;\;\;(1 \leqslant m \leqslant 10, m \leqslant n \leqslant 10) $

  (3)

  Where T is the sign function, indicating the entropy change resulting from different types of atomic interactions^[16-20]; S represents the solvent accessible surface area of atom (SASA), i.e., surface area formed by a hydrateprobe spherecally rolling at the surface of this atom^[21]; a is the hydrophobic constants^[22].

  Molecular docking was performed by AutoDock 4.2 software package^[23]. Using simulated annealing and genetic algorithm to find the best junction of close position of the receptor and the ligand, by computing free energy of the semi-empirical method, we evaluate the matching in receptor and the ligand, and using functional form (type (4))^[24] to calculate.

  $ \Delta G = \Delta {G_{vdw}}\sum {\left( {\frac{{{A_{ij}}}}{{r_{ij}^{12}}} - \frac{{Bij}}{{r_{ij}^6}}} \right)} + \Delta G_{H - bond}\sum\limits_{i, j} E \left( t \right)\left( {\frac{{\mathop C\nolimits_{ij} }}{{\mathop r\nolimits_{ij}^{10} }}} \right) + \Delta {G_{ele}}\sum\limits_{i, j} {\frac{{{q_{ij}}}}{{\varepsilon \left( {{r_{ij}}} \right){r_{ij}}}}} + \Delta {G_{tor}}{N_{tor}} + \Delta {G_{sol}}\sum\limits_{i, j} {\left( {{S_i}{V_j} + {S_j}{V_i}} \right){e^{\left( {{\raise0.7ex\hbox{${r_{ij}^2}$} \!\mathord{\left/ {\vphantom {{r_{ij}^2} {2{\delta ^2}}}}\right.} \!\lower0.7ex\hbox{${2{\delta ^2}}$}}} \right)}}} $

  (4)

  Where ΔG_vdw, ΔG_H-bond, ΔG_ele, ΔG_tor and ΔG_sol are the semi-empirical parameters obtained by fitting.

  3. RESULTS AND DISCUSSION

  3.1 Selection and partition of the data

  28 Thiazolidine-4-carboxylic acid derivatives with inhibitory influenza A virus were obtained from references^[25]. IC₅₀ values were measured spectrofluoro-metrically using 2΄-(4-methylumbelliferyl)-a-D-acetyl neuraminic acid (MUNANA) as substrate for neuraminidase to yield a fluorescent product which was quantified^[25]. From structurally diverse molecules possessing activities of a wide range, 28 NIs were divided into the training set with 22 samples and the test set containing 6 samples. The common structure is shown in Fig. 1, and the structures and predicted activity values of 28 thiazolidine-4-carboxylic acid derivatives are listed in Table 1.

  Figure 1

  

  Figure 1.  Skeleton of 28 thiazolidine-4-carboxylic acid derivatives

  DownLoad: Full-Size Img PowerPoint

  Table 1

  

  Table 1.  Structures and Predicted Activities of 28 Thiazolidine-4-carboxylic Acid Derivatives

  DownLoad: CSV

  Compound	R1	R2	Exp. pIC50	Pred. pIC50	Residual
  1		H	4.6716	4.6670	0.0046
  *2		H	4.6946	4.9002	–0.2056
  3		H	4.7423	4.6383	0.1040
  *4		H	4.6308	4.5349	0.0959
  5		H	4.6478	4.6692	–0.0214
  6		H	4.9101	4.8693	0.0408
  7		H	4.3655	4.4668	–0.1013
  8		ClCH2CO-	5.1232	5.1261	–0.0029
  		ClCH2CO-	5.2336	5.2141	0.0195
  *10		ClCH2CO-	4.9706	4.9139	0.0567
  11		ClCH2CO-	5.0635	5.1343	–0.0708
  12		ClCH2CO-	5.1163	5.1101	0.0062
  13		ClCH2CO-	5.1013	4.9952	0.1061
  14		ClCH2CO-	4.8894	4.8974	–0.0080
  *15		PhCH2CO-	5.9172	5.8849	0.0323
  16		PhCH2CO-	6.1871	5.7805	0.4066
  17		PhCH2CO-	5.7167	5.9931	–0.2764
  18		PhCH2CO-	5.6073	5.7234	–0.1161
  19		PhCH2CO-	5.7282	5.7648	–0.0366
  *20		PhCH2CO-	5.7905	5.8312	–0.0407
  21		PhCH2CO-	5.5391	5.5700	–0.0309
  22		NH2CH2CO-	6.2757	6.3593	–0.0836
  23		NH2CH2CO-	6.6778	6.6281	0.0497
  *24		NH2CH2CO-	6.5528	6.1802	0.3726
  25		NH2CH2CO-	6.0915	6.1459	–0.0544
  26		NH2CH2CO-	5.9914	6.3027	–0.3113
  27		NH2CH2CO-	6.8539	6.6821	0.1718
  28		NH2CH2CO-	6.0088	6.1156	–0.1068
  *Samples in the test set

  Molecular steric structures were firstly constructed by Chemoffice 8.0, and then optimized at the AM1 level by MOPAC half-experience quantum chemistry software in Chem3D. Then net electric charge of atoms was calculated in single-point form by Mulliken methods. After the above two items were input respectively into forms of Descartes coordinates and net electric charge amounts, 3D-HoVAIF descriptors were produced by applying 3D-HoVAIF. EXE is an applied program written in C language. The ultimate vectors for 28 thiazolidine-4-carboxylic acid derivatives involve 105 non-zero items. The interpreted unification of the 3D-HoVAIF method may lead to some information overlap among these different descriptors. To address the above mentioned problems, step wise multiple regression (SMR) method was employed which used the SPSS 13.0 software to select variables; Multiple linear regression (MLR) and partial least squares (PLS) ways were applied to construct the model according to the values of Fisher prominent test by SMR. The obtained original variable matrix by SMR was then subjected to a PLS regression modeling or MLR regression modeling and the optimal model was determined when cross-validation correlative coefficients (R_CV²) in leave-one-out cross-validation (LOO-CV) achieved the maximum value, and the relative statistics to these models were presented. QSAR models were validated by leave-one-out cross-validation (LOO-CV) and external prediction by test set R_test² and Q_ext^2[26]. Meanwhile, the models were validated by different modeling methods.

  $ {Q_{ext}}^2 = 1 - \frac{{\sum\limits_{i = 1}^{test} {{{\left( {{y_i} - {{\hat y}_i}} \right)}^2}} }}{{\sum\limits_{i = 1}^{test} {{{\left( {{y_i} - {{\bar y}_{tr}}} \right)}^2}} }} $

  (5)

  In Eq. 5, both $ {y_i} $ and $ {\hat y_i} $ are the observed and calculated values of the test dataset, and $ {y_{tr}} $ is the mean value of observed values of the training dataset.

  3.2 Model building

  In this study, two models, PLS and MLR, were obtained. PLS was widely used in analytical, physical, and medicinal chemistry as a data analyzing method. It has many advantages over ordinary MLR. For instance, it can avoid harmful effects in modeling due to multi-collinearity, and it particularly fits for regressing when the number of variables is less than that of samples, etc. In QSAR studies, variables, as few as possible, should be included in a model. Not all the structural descriptors relevant to biological activities will be easily interpreted, so those redundant descriptors should be eliminated in order to promote its robustness and predictive capability especially when the number of variables is very large. The correlation coefficients (R²) of established PLS model was 0.964, and the cross-validated correlation coefficients (Q²) of PLS model was 0.935. Furthermore, the correlation coefficient for the test set (R_test²) was 0.911 and the cross-validated correlation coefficient for the test set (Q_ext²) was 0.644, respectively.

  However, PLS model has so many advantages but MLR model has also many traits. Linear Regression Multiple (MLR) is a more classic modeling method, which can optimize the activity of the lead compounds. The assumption of the MLR method is that when the molecular structure changes, the biological activity of the molecules will change, and the fundamental cause of this change is closely related to the physical parameters. In addition, another obvious advantage of it is that it can obtain a causal model, and has a clear physical meaning. In this paper, the results of MLR were better than those of PLS. The results obtained by MLR were as follows:

  $ \begin{aligned} & \mathrm{p} I C_{50}=2.8450 .000 \times H_{3-5}-0.841 \times E_{1-8}-0.364 \times \\ & E_{1-3}-0.002 \times S_{2-5}+0.000 \times H_{1-5}+0.000 \times \\ & {H_{5-5}}^-200.684 \times S_{2-10} \\ & N=21, R_{\mathrm{cum}}=0.992, S D=0.107, F=119.322, \\ & {Q_{\mathrm{cv}}}^2=0.947, S D_{\mathrm{cv}}=0.199, F_{\mathrm{cv}}=33.138, \\ & t_1=-7.903, t_2=-9.618, t_3=-3.235, t_4=3.397, \\ & t_5=1.960, t_6=1.618, t_7=-0.653 \end{aligned} $

  N for the training samples, R_cum as the correlation coefficient, SD is the standard deviation, F is F test values, Q_CV², SD_CV, F_CV and t_i for the cross-validated correlation coefficients, standard deviation, F test value and t test value of each variable.

  In the model, H_3-5 is hydrophobic function between C of sp² hybridization and N of sp³ hybridization, E_1–8 is electrostatic function between H and O of sp³ hybridization, E_1-3 is electrostatic function between H and C of sp² hybridization, S_2–5 is steric function between C of sp³ hybridization and N of sp³ hybridization, H_1–5 is hydrophobic function between H and N of sp³ hybridization, H_5–5 is hydrophobic function of N of sp³ hybridization, S_2-10 is steric function between C of sp³ hybridization and C_l.

  3.3 Model analysis

  All the parameters in the model of Table 2 were taken from the Ref. [7], indicating that the obtained model of 3D-HoVAIF is superior to other models, so it is robust with good exterior predictive capabilities. Although the method is old and troublesome, the result is good. Hence, for the model, the method of 3D-HoVAIF was chosen at last.

  Table 2

  

  Table 2.  Statistical Results and Relative Contributions of the Models

  DownLoad: CSV

  Methods	R2	SE	Q2	SECV	Rtest2	Qext2	A	F
  HQSAR	0.994	0.066	0.951	0.185	0.879	0.867	6	-
  Topomer CoMFA	0.978	0.246	0.919	-	0.912	0.884	7	-
  COMSIA	0.996	0.049	0.820	0.434	0.953	0.899	3	786.865
  HoVAIF	0.984	0.107	0.947	0.199	-	0.967	7	119.322
  R2: squared multiple correlation coefficients; SE: standard error of estimate; Q2: squared cross-validated correlation coefficient; SEcv: cross-validated standard error of estimate; A: components; F: Fisher statistic; Rtest2 and Qext2 are R2 and Q2 of the test set, respectively.

  From the above results, it can be seen that the results of PLS are worse than those of MLR, so the way of MLR was obtained to predict the test set. At the same time, the predicted activities of training and test sets compounds were achieved. And then the scatter plot between calculated and predicted activities of the training and test sets compounds was given in Fig. 2. As shown in Fig. 2, calculated activities are in a linear relationship with predicted activities of the samples. It told that all samples were almost uniformly distributed around the diagonal, and no obviously exceptional point was selected. Q_ext² was 0.967, so the results of internal and external inspection showed the model was accurate and stable. Thereby, the obtained QSAR model with good exterior predictive capability was robust. The adopted QSAR model can be used to design and screen new NIs.

  Figure 2

  

  Figure 2.  Scatter plot between the observed and predicted activity of training and test sets of 28 NAIs

  DownLoad: Full-Size Img PowerPoint

  3.4 Drug molecular design

  From the above regression equation, the pIC₅₀ increases with the decrease of E_1–3, E_1-8, S_2–5 and S_2-10 values, which are to enhance the activities of drug molecules. So the highest activity compound 27 of 28 thiazolidine-4-carboxylic acid derivatives which have been synthesized and evaluated as neuraminidase inhibitors was chosen for optimization, and then it would get higher activity compounds. In accord with compound 27 as the template for the design of new drug molecules, newly designed molecules and predictive activity are shown in Table 3. By comparing the newly designed compound 1 with compound 27 of the original template molecule, the substitution with -COOH group of the former showed better activity than the substitution with -OH group of the template molecule. On account of Π=Π conjugate, the biological activity of compounds increases. By comparing the newly designed compound 6 with compound 27 of the original template molecule, the (CH₃)₂CH- of compound 6 replaced CH₃O- of the template molecule, which can increase the biological activity by decreasing the electrostatic interaction.

  Table 3

  

  Table 3.  Newly Designed Molecular Structures and Predictive Activity

  DownLoad: CSV

  Compound	R1	R2	Pred.pIC50
  1		NH2CH2CO-	6.696
  2		NH2CH2CO-	6.390
  3		NH2CH2CO-	6.489
  4		NH2CH2CO-	6.572
  5		NH2CH2CO-	6.611
  6		NH2CH2CO-	6.747

  3.5 Docking study

  Docking was performed using AutoDock software package version 4.2. 3D co-crystallized structure of neuraminidase was taken from the Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank (PDB ID: 4B7R). Processing of the protein was done by adding hydrogen atoms to assign appropriate ionization states to both acidic and basic amino acid residues and removing the remaining water molecules from protein and adding charges. In this experiment, the No. 27 compound with the highest activity of template and No. 6 compound with the highest activity of newly designed compounds were adopted by molecular docking, respectively, docking lattice was 40 × 40 × 40, and grid point spacing was 0.375 Angstroms, coordinates of the central grid point of maps for No. 27 template compound was (23.282, 21.870, –42.954), while coordinates of the central grid point of maps for newly designed compound 6 was (40.696, –1.808, –5.945), Lamarckian Genetic AlGotithm (LGA) was used as ligand conformation search process, and the other parameters were by default.

  The docked conformations showed that all ligands bind to the active residues in the predefined hydrophobic binding pocket. As can be seen from Fig. 3, the location of the docked ligands agreed well with that of the docked reference ligand. For Fig. 3a, the reference ligand formed hydrogen bonding (H···bond) with ASP151, GLU278, TYR402, ARG293 and ARG118 in the binding sites. The docked reference ligand was found to have H···bond interactions: NH of sp³ hybridization of the ligand and O group of ASP151 (hydrogen bond distance 2.1 nm), OH of sp³ hybridization of ligand and O group of GLU278 (hydrogen bond distance 2.1 nm), C=O of sp² hybridization of ligand and NH groups of ARG118 and ARG368 (hydrogen bond distances 2.1 and 1.7 nm, respectively), OH of sp³ hybridization of the ligand and OH group of TYR402 (hydrogen bond distance 1.9 nm) and OH of sp³ hybridization of the ligand and NH group of ARG293 (hydrogen bond distance 2.1 nm).

  Figure 3

  

  Figure 3.  Docking interaction pattern of 4B7R active residues with ligands

  (a) Compound 27 and 4B7R active residues, (b) New compound 6 and 4B7R active residues

  DownLoad: Full-Size Img PowerPoint

  However, for Fig. 3b, the reference ligand formed hydrogen bonding (H···bond) with GLU277, ARG368 and ARG293 in the binding sites. The docked reference ligand was found to have H···bond interaction between NH of sp³ hybridization of the ligand and O group of GLU277 (hydrogen bond distance 2.3 nm), C=O of sp² hybridization of ligand and NH₂ group of ARG368 (hydrogen bond distance 2.3 nm) and OH of sp³ hybridization of ligand and NH₂ group of ARG293 (hydrogen bond distance 2.3 nm).

  From molecular docking model, in the docking process of ligand and receptor, the formation of hydrogen bonds between them determined the ligands in live. The positions of the cavity have an important role. Hydrogen bonding between drug molecules and biological macromolecule receptors is a common way of bonding, which can increase the drug activity and make the combination between drug molecules more stable. The docking results agreed well with the observed biological activity data, which showed that these docking conformations are desirable to analyze the drug models.

  4. CONCLUSION

  28 Thiazolidine-4-carboxylic acid derivatives are a kind of anti-neuraminidase drug having great development prospects and good exterior predictive capability by using 3D-HoVAIF way. In this paper, 3D-QSAR model was set up by means of 3D-HoVAIF and MLR, finally achieving good results. It is illustrated that the obtained 3D-QSAR is reliable for describing chemical structure and biological activity of the drug molecules, and then molecular docking approach was employed to study the relationship between drug ligand and macromolecular receptor. Ultimately, the docking results indicate that the ligands would form hydrogen bonding interacttions with GLU, TYR, ARG and ASP of the protein receptor, generally. So, the adopted QSAR model can help to design and screen new compounds to obtain new NAIs with high activities.

  
  
• 1. [1]

     Chen, Y. Q.; Wohlbold, T. J.; Zheng, N. Y. Influenza infection in humans induces broadly cross-reactive and protective neuraminidase-reactive antibodies. Cell 2018, 173, 417−429. doi: 10.1016/j.cell.2018.03.030

  2. [2]

     Wang, X. L.; Jiang, H.; Wu, P. Epidemiology of avian influenza A H7N9 virus in human beings across five epidemics in mainland China, 2013-17: an epidemiological study of laboratory-confirmed case series. Lancet Infect Dis. 2017, 17, 822−832. doi: 10.1016/S1473-3099(17)30323-7

  3. [3]

     Zhao, H. Y.; Chen, Z. L. Screening of neuraminidase inhibitors from traditional Chinese medicines by integrating capillary electrophoresis with immobilized enzyme microreactor. J. Chromatogr. A 2014, 1340, 139−145. doi: 10.1016/j.chroma.2014.03.028

  4. [4]

     Anne Moscona, M. D. Neuraminidase inhibitors for influenza. New Engl. J. Med. 2005, 353, 1363−1373. doi: 10.1056/NEJMra050740

  5. [5]

     Chen, B. Y.; Zhang, T.; Bond, T. Development of quantitative structure activity relationship (QSAR) model for disinfection byproduct (DBP) research: a review of methods and resources. J. Hazard. Mater. 2015, 299, 260−279. doi: 10.1016/j.jhazmat.2015.06.054

  6. [6]

     Bhole, R. P.; Bhusari, K. P. Synthesis, antihypertensive activity, and 3D-QSAR studies of some new p-hydroxybenzohydrazide derivatives. Arch. Pharm. 2011, 344, 119−134. doi: 10.1002/ardp.201000008

  7. [7]

     Sun, J. Y.; Wang, J. C.; Mei, H. QSAR and pharmacophore studies of thiazolidine-4-carboxylic acid derivatives as novel influenza neuraminidase inhibitors using HQSAR, Topomer CoMFA and CoMSIA. Chin. J. Struc. Chem. 2013, 32, 744−750.

  8. [8]

     Van Damme, S.; Bultinck, P. 3D QSAR based on conceptual DFT molecular fields: antituberculotic activity. J. Mol. Struc.-Theochem. 2010, 943, 83−89. doi: 10.1016/j.theochem.2009.10.031

  9. [9]

     Tong, J. B.; Liu, S. L. Three-dimensional holographic vector of atomic interaction field applied in QSAR of anti-HIV HEPT analogues. QSAR Comb. Sci. 2008, 27, 330−337. doi: 10.1002/qsar.200710076

  10. [10]

      Tong, J. B.; Chen, Y.; Liu, S. L. QSAR studies of antituberculosis drug using three-dimensional structure descriptors. Med. Chem. Res. 2013, 22, 4946−4952. doi: 10.1007/s00044-013-0502-y

  11. [11]

      Tong, J. B.; Che, T.; Li, Y. F. A descriptor of amino acids: SVRG and its application to peptide quantitative structure-activity relationship. SAR QSAR Environ. Res. 2011, 22, 611−620. doi: 10.1080/1062936X.2011.604099

  12. [12]

      Levitt, M. Protein folding by restrained energy minimization and molecular dynamics. J. Mol. Biol. 1983, 170, 723−764. doi: 10.1016/S0022-2836(83)80129-6

  13. [13]

      Levitt, M.; Perutz, M. F. Aromatic rings act as hydrogen bond acceptors. J. Mol. Biol. 1988, 201, 751−754. doi: 10.1016/0022-2836(88)90471-8

  14. [14]

      Hahn, M. Receptor surface models. 1. Definition and construction. J. Med. Chem. 1995, 38, 2080−2090. doi: 10.1021/jm00012a007

  15. [15]

      Kellogg, G. E.; Semus, S. F.; Abraham, D. J. HINT: a new method of empirical hydrophobic field calculation for CoMFA. J. Comput. Aid. Mol. Des. 1991, 5, 545−552. doi: 10.1007/BF00135313

  16. [16]

      Wireko, F. C.; Kellogg, G. E.; Abraham, D. J. Allosteric modifiers of hemoglobin. 2. Crystallographically determined binding sites and hydrophobic binding/interaction analysis of novel hemoglobin oxygen effectors. J. Med. Chem. 1991, 34, 758−767. doi: 10.1021/jm00106a042

  17. [17]

      Tong, J. B.; Zhong, L.; Zhao, X. Quantitative structure-activity relationship studies of diarylpyrimidine derivatives as anti-HIV drugs using new three-dimensional structure descriptors. Med. Chem. Res. 2014, 23, 1634−1642. doi: 10.1007/s00044-013-0770-6

  18. [18]

      Kellogg, G. E.; Abraham, D. J. Key, lock, and locksmith: complementary hydropathic map predictions of drug structure from a known receptor-receptor structure from known drugs. J. Mol. Graph. 1992, 10, 212−217. doi: 10.1016/0263-7855(92)80070-T

  19. [19]

      Tong, J. B.; Chen, Y.; Liu, S. L. A novel 3D molecular structural characterization method applied in QSAR study. Rev. Chim.-Bucharest. 2013, 64, 707−712.

  20. [20]

      Li, Z. H.; Chen, G.; Chen, Z. T. Three-dimensional holographic vector of atomic interaction field (3D-HoVAIF) for the QSPR/QSAR of polychlorinated naphthalenes. Chin. J. Struct. Chem. 2012, 31, 345−352.

  21. [21]

      Shu, M.; Zhang, Y. R.; Tian, F. F. Molecular docking and 3D-QSAR research of biphenyl carboxylic acid MMP-3 inhibitors. Chin. J. Struct. Chem. 2012, 31, 443−451.

  22. [22]

      Liao, L. M.; Huang, X.; Li, J. F. Structural characterization and acute toxicity simulation for nitroaromatic compounds. Chin. J. Struct. Chem. 2016, 35, 449−456.

  23. [23]

      Huey, R.; Morris, G. M.; Olson, A. J. A semiempirical free energy force field with charge-based desolvation. J. Comput. Chem. 2007, 28, 1145−1152. doi: 10.1002/jcc.20634

  24. [24]

      Liu, Y.; Jing, F. B.; Xu, Y. Y. Design, synthesis and biological activity of thiazolidine-4-carboxylic acid derivatives as novel influenza neuraminidase inhibitors. Bioorgan. Med. Chem. 2011, 19, 2342−2348. doi: 10.1016/j.bmc.2011.02.019

  25. [25]

      Golbraikh, Alexander; Tropsha, Alexander. Beware of q²! J. Mol. Graph. Model. 2002, 20, 269−276. doi: 10.1016/S1093-3263(01)00123-1

  26. [26]

      Tong, J. B.; Wang, Y.; Lei, S. 3D-QSAR and docking studies of 1,3,4-thiazolidinone derivatives using R-group search and surflex-dock. Chin. J. Struct. Chem. 2019, 38, 464−475.

• 

  Figure 1  Skeleton of 28 thiazolidine-4-carboxylic acid derivatives

  下载: 全尺寸图片 幻灯片
  

  Figure 2  Scatter plot between the observed and predicted activity of training and test sets of 28 NAIs

  下载: 全尺寸图片 幻灯片
  

  Figure 3  Docking interaction pattern of 4B7R active residues with ligands

  (a) Compound 27 and 4B7R active residues, (b) New compound 6 and 4B7R active residues

  下载: 全尺寸图片 幻灯片

  Table 1.  Structures and Predicted Activities of 28 Thiazolidine-4-carboxylic Acid Derivatives

  Compound	R1	R2	Exp. pIC50	Pred. pIC50	Residual
  1		H	4.6716	4.6670	0.0046
  *2		H	4.6946	4.9002	–0.2056
  3		H	4.7423	4.6383	0.1040
  *4		H	4.6308	4.5349	0.0959
  5		H	4.6478	4.6692	–0.0214
  6		H	4.9101	4.8693	0.0408
  7		H	4.3655	4.4668	–0.1013
  8		ClCH2CO-	5.1232	5.1261	–0.0029
  		ClCH2CO-	5.2336	5.2141	0.0195
  *10		ClCH2CO-	4.9706	4.9139	0.0567
  11		ClCH2CO-	5.0635	5.1343	–0.0708
  12		ClCH2CO-	5.1163	5.1101	0.0062
  13		ClCH2CO-	5.1013	4.9952	0.1061
  14		ClCH2CO-	4.8894	4.8974	–0.0080
  *15		PhCH2CO-	5.9172	5.8849	0.0323
  16		PhCH2CO-	6.1871	5.7805	0.4066
  17		PhCH2CO-	5.7167	5.9931	–0.2764
  18		PhCH2CO-	5.6073	5.7234	–0.1161
  19		PhCH2CO-	5.7282	5.7648	–0.0366
  *20		PhCH2CO-	5.7905	5.8312	–0.0407
  21		PhCH2CO-	5.5391	5.5700	–0.0309
  22		NH2CH2CO-	6.2757	6.3593	–0.0836
  23		NH2CH2CO-	6.6778	6.6281	0.0497
  *24		NH2CH2CO-	6.5528	6.1802	0.3726
  25		NH2CH2CO-	6.0915	6.1459	–0.0544
  26		NH2CH2CO-	5.9914	6.3027	–0.3113
  27		NH2CH2CO-	6.8539	6.6821	0.1718
  28		NH2CH2CO-	6.0088	6.1156	–0.1068
  *Samples in the test set

  下载: 导出CSV

  Table 2.  Statistical Results and Relative Contributions of the Models

  Methods	R2	SE	Q2	SECV	Rtest2	Qext2	A	F
  HQSAR	0.994	0.066	0.951	0.185	0.879	0.867	6	-
  Topomer CoMFA	0.978	0.246	0.919	-	0.912	0.884	7	-
  COMSIA	0.996	0.049	0.820	0.434	0.953	0.899	3	786.865
  HoVAIF	0.984	0.107	0.947	0.199	-	0.967	7	119.322
  R2: squared multiple correlation coefficients; SE: standard error of estimate; Q2: squared cross-validated correlation coefficient; SEcv: cross-validated standard error of estimate; A: components; F: Fisher statistic; Rtest2 and Qext2 are R2 and Q2 of the test set, respectively.

  下载: 导出CSV

  Table 3.  Newly Designed Molecular Structures and Predictive Activity

  Compound	R1	R2	Pred.pIC50
  1		NH2CH2CO-	6.696
  2		NH2CH2CO-	6.390
  3		NH2CH2CO-	6.489
  4		NH2CH2CO-	6.572
  5		NH2CH2CO-	6.611
  6		NH2CH2CO-	6.747

  下载: 导出CSV
•