Design, virtual screening and in silico QSPR modeling for the development of new thiosemicarbazone-based complexes

Chia sẻ: _ _ | Ngày: | Loại File: PDF | Số trang:9

Thêm vào BST

Báo xấu

8
lượt xem 2
download

Download Vui lòng tải xuống để xem tài liệu đầy đủ

This modeling is performed on an experimental data set of complexes, where the metal ions of these complexes include transition ion metals and lanthanide ion metals. We use these models to develop a series of new thiosemicarbazone and their complexes; simultaneously, the complexes are worked out the stability constants from the novel models.

Chủ đề:

Bình luận(0) Đăng nhập để gửi bình luận!

Lưu

Nội dung Text: Design, virtual screening and in silico QSPR modeling for the development of new thiosemicarbazone-based complexes

Cite this paper: Vietnam J. Chem., 2023, 61(S1), 8-16 Research article DOI: 10.1002/vjch.202200203 Design, virtual screening and in silico QSPR modeling for the development of new thiosemicarbazone-based complexes Nguyen Minh Quang1, Huynh Ngoc Chau1, Tran Thai Hoa2*, Vu Thi Bao Ngoc4, Pham Van Tat3* 1 Faculty of Chemical Engineering, Industrial University of Ho Chi Minh City, 12 Nguyen Van Bao, Go Vap, Ho Chi Minh City 70000, Viet Nam 2 Faculty of Chemistry, Hue University of Sciences, Hue University, 77 Nguyen Hue, Hue City 49100, Viet Nam 3 Department of Sciences and Journal Management, Hoa Sen University, 8 Nguyen Van Trang, Dist. 1, Ho Chi Minh City 70000, Viet Nam 4 Faculty of Chemistry and Environment, University of Dalat, 01 Phu Dong Thien Vuong, Da Lat City 66000, Viet Nam Submitted November 16, 2022; Revised January 16, 2023; Accepted February 14, 2023 Abstract Eighteen new thiosemicarbazone ligands and 30 new ligand-based complexes were developed from quantitative structure-property relationships (QSPR) methods. Stability constants (log12) of complexes were calculated on QSPR models that were built by methods of multivariate linear regression (MLR) and artificial neural network (ANN). Six descriptors, including dipole, 5C, 4N, fw, xc3, and ka1 were discovered in the best QSPRMLR model with the good statistical criteria: R2train = 0.892, Q2CV = 0.845, and SE = 0.900. Besides, the ANN model with architecture I(6)-HL(3)- O(1) was built from the descriptors of the MLR model with excellent results as R2train = 0.958, Q2CV = 0.966, and Q2test = 0.980. Also, the models were externally validated on the other experimental dataset. Consequently, the resulting QSPR models could be applied to develop new complexes for chemically related fields. Keywords. Machine learning, MLR, QSPR, stability constants log12, thiosemicarbazone. 1. INTRODUCTION thiosemicarbazones leads to a variety of practical applications. Some have therapeutic activity used as Recently, a new theoretical method is emerging as a antivirals and antibiotics;[2] some have been used as potential method for developing new derivatives. The anticancer drugs.[2] Besides, thiosemicarbazone has method is based on the relationship between a certain also been used as a potential reagent in the formation quantitative value that characterizes the properties of of complexes,[3] which is understandable because the research object and its structure, and the method their structure contains nitrogen and sulfur donors, so is called the quantitative structure-properties the ability to form complexes is very easy. Moreover, relationship (QSPR). This method came from a these complexes also have properties and activities classic study, namely the quantitative structure- similar to those of thiosemicarbazone mentioned activity relationship (QSAR), activity was replaced above.[2,3] This creates excitement in the research and by the property.[1] This complex method covers steps new development of these derivatives. The easy such as data mining, filtration, molecular or quantum complexation is a prerequisite for application in mechanism computation for optimizing structures, analytical chemistry by using the UV-Vis technique. and the construction of statistical and mathematically It is a cheap, quick, easy operation, and highly rigorous regression equations.[1] Up to now, many effective technique. In fact, many works have been works have successfully used this method to develop published about thiosemicarbazones used as a reagent new derivatives. in analytical techniques.[3] Meanwhile, thiosemicarbazone derivatives are Up to now, there are a few QSPR studies on the also known as organosulfur and organonitrogen M:L complex of a thiosemicarbazone ligand (L) with compounds and the structural peculiarity of a metal ion (M) in an aqueous solution.[4,5] 8 Wiley Online Library © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Tran Thai Hoa et al. Meanwhile, the difference between these studies is and their constant stability values. Therefore, the that the models are developed with different data sets standard QSPR model is expressed by the following and different methods. Besides, in the QSAR/QSPR equation:[1] research, popular methods such as multivariate linear logpq = f(descriptors) (1) regression (MLR), support vector regression (SVR), Multivariable linear regression (MLR) and and artificial neural network (ANN) are often used or artificial neural network (ANN) methods are selected can be combined with other intelligent algorithms.[4,5] to build the models, and the development of models However, the studies on the M:2L complex of two complies with OECD rules and the instructions of thiosemicarbazone ligands (L) with a metal ion (M) Golbraikh and Tropsha.[1] This modeling is in an aqueous solution are still very limited and these performed on an experimental data set of complexes, studies also came from our works.[6] Furthermore, it where the metal ions of these complexes include should be noted that in QSAR/QSPR studies the data transition ion metals and lanthanide ion metals. We space is so large that it is difficult to have a model that use these models to develop a series of new represents the entire structure of the derivative. It thiosemicarbazone and their complexes; means that each data group is limited to a group of simultaneously, the complexes are worked out the structural features of the study derivatives. Therefore, stability constants from the novel models. the aim of this study is to develop a QSPR model of the M:2L complex on a pool of carefully screened 2. METHODOLOGY data and characterization of the structure derivatives group. 2.1. The collection of data and data sets Thereupon, the constant stability values are selected as a characteristic quantity of the complex The complex of the ML2 type was selected in this properties in order to develop new thiosemicarbazone work. The complexes are generated between one and their complexes. To do this, we use the QSPR metal ion (M) and two thiosemicarbazones (L). modeling method to build predictive models between Skeletons of the ligand and complex are displayed in the structures of thiosemicarbazone-based complexes figure 1. (a) (b) Figure 1: (a) The sketch of research ligand; (b) Complex between Cu2+/Ni2+ and 2-acetylpyridine-4-methyl-3-thiosemicarbazone[7] Data collection is the first stage of the series step models, and they characterize the structure of the in the QSAR/QSPR modeling. First, big data was research object. They included topological and collected from experimental results; then, the physicochemical parameters for the molecular clustering method was used to split it into some data structure (0-3D descriptors) and were calculated on subsets.[1] In this work, two subsets were used to build the QSARIS tool, while the quantum parameters were the models. The first subset is used to build the calculated from the optimal structure by using the models consisting of 21 ligands and the 62 stability quantum computation method with semi-empirical constants values (logβ12) of the metal- methods PM7 and PM7/sparkle on MOPAC2016 [8]. thiosemicarbazone complexes (table 1). The second In this study, 230 molecular descriptors for each of subset is used for external evaluation of the built the complexes were calculated by QSARIS and 24 models (table 5) including 10 thiosemicarbazone quantum and partial charge descriptors were ligands and the 20 stability constants values (logβ12) generated from the results of above-mentioned of the ligand-based complexes. quantum computation method. After that, the descriptors were carefully screened by removing the 2.2. Calculation of descriptors descriptors with more than 70% similar values. Collecting all data, including the 104 descriptors and Descriptors are the variables of the QSAR/QSPR the logarithm of stability constants of each © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 9
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Design, virtual screening and in silico… thiosemicarbazone-based complexes, forms a data and k is the number of variables of the model. matrix used to build QSPR models. The construction of QSPRMLR models was carried out by a step-by-step regression approach based on 2.3. QSPR Construction the Regression technique.[9] It is an add-in function of the MS-EXCEL platform. Furthermore, the models 2.3.1. QSPRMLR model were cross-validation (CV) by the leave-one-out (LOO) method[1,9] via the Q2CV criteria. The statistical The multivariate linear regression (MLR) method is a parameters controlled the quality of the QSPRMLR popular technique used the QSAR/QSPR modeling; model like adjust variables (R2adj), coefficient of Generally, the MLR model is represented by the determination (R2train), standard error (SE), Fischer's following expression:[1,9] k value (Fstat), and predicted residual sum of squares y = 0 +  i xi (2) (PRESS). In addition, MLR models were externally i =1 validated on an independent dataset using the Q2ext where 0 is the constant, i is the regression coefficient, criteria. Table 1: The minimum (logβ12a) and maximum (logβ12b) stability constants values of the experimental data set Thiosemicarbazone ligand Metal Number of No logβ12a logβ1b Ref. R1 R2 R3 R4 ions complexes, n 1 H -CH3 -CH3 -C5H4N Ni2+ 1 11.919 11.919 [10] 2 H H -CH3 -C7H5N2 Co2+ 1 16.694 16.694 [11] 3 H H -CH3 -C5H4N Hg2+ 1 12.875 12.875 [7] 4 H H H -C5H4N Fe2+ 1 18.180 18.180 [12] 5 H H -CH3 -C5H4N Fe2+ 1 19.230 19.230 [12] 6 H H -C5H4N -C5H4N Zn2+ 1 10.370 10.370 [12] 7 H H -CH3 -C5H4N La3+ 4 13.500 14.040 [13] 8 H H -CH3 -C5H4N Pr3+ 4 13.830 14.530 [13] 9 H H -CH3 -C5H4N Nd3+ 4 14.180 14.990 [13] 10 H H H -C10H6OH Pr3+ 4 12.940 15.060 [14] 11 H H H -C10H6OH Nd3+ 4 13.500 15.030 [14] 12 H H H -C10H6OH Sm3+ 4 13.800 15.580 [14] 13 H H H -C10H6OH Eu3+ 4 14.210 15.430 [14] 14 H H H -C10H6OH Gd3+ 4 14.380 15.550 [14] 15 H H H -C10H6OH Y3+ 4 14.350 15.700 [14] 16 H H H -C10H6OH Tb3+ 4 14.660 15.970 [14] 17 H H H -C10H6OH Dy3+ 4 14.950 16.410 [14] 18 H H -CH3 -C6H4OH Mn2+ 3 7.2400 8.5200 [15] 19 H H -CH3 -C6H4OH Cd2+ 3 7.9000 9.7900 [15] 20 H H -CH3 -C6H4OH Ni2+ 3 8.6300 10.380 [15] 21 H H -CH3 -C6H4OH Cu2+ 3 9.8100 11.960 [15] 2.3.2. QSPRANN modeling MLR model, an output layer (n) is the stability constant (log12), and the number of hidden neurons Generally, an artificial neural network (ANN) is a (m) is determined by neurons on the input and output machine learning method and it is widely used in the layer. modeling technique and applied in many fields.[16] The process of training ANN covered two stages. This study used to be training deep neural networks At the start, the number of neurons in the hidden layer (DNN), namely multi-layer perceptron (MLP) type was found by the Neural Designer package, and this and Levenberg-Marquardt back-propagation step aims to record the architecture of ANN with good algorithm,[17] so the architecture of ANN includes predictability. After that, the neural networks were three layers I(k)-HL(m)-O(n) that is an input layer, externally evaluated on an independent dataset to one hidden layer, and an output layer. The number of select the best QSPRANN model, and the second stage neurons in the input layer (k) is the descriptor of the was operated on Matlab. Furthermore, the © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 10
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Tran Thai Hoa et al. performance of the DNN modeling used the primary Furthermore, comparing the quality of the models transfer functions of neural networks, which are the was operated by using the quantity of the mean hyperbolic tangent-sigmoid and log-sigmoid absolute value of the relative error MARE (%). It is functions.[17] calculated on the formula (4): The ANN models were trained in the initial 1 log 12,exp − log 12,pred dataset and randomly divided into training, MARE,% = .100 (4) N log 12,exp validation, and testing subsets. The quality of the QSPRANN models has commanded corroding to three where N is the number of samples; β12,exp is the coefficients regression such as R2train, Q2CV, and Q2test. experimental stability constants; β12,pred is the Besides, the training of ANN models will stop by the predicted stability constants. minima of the mean squared error (MSE) value, and lowers values are better. MSE is the average squared 3. RESULTS AND DISCUSSION difference between training outputs (t) and experimental outputs (e).[17] 3.1. Constructing QSPRMLR models 2.4. The contribution of the variables and the As stated, the dataset for constructing QSPRMLR indicators for comparison of the models models includes the 62 observation values of thiosemicarbazone-based complexes, and it was cut This work used the quantity of the average into a training subset of 50 values (approximation of contribution percentage (ACPxk,i) to evaluate the 80%) and a CV subset of 12 values (approximation of influence of the variables in the models, and its 20%).[1] The quality of the models depends on the mathematical expression is described by the formula criteria of statistics to classify the goodness of (3): models.[1] The results of QSPRMLR models were built entirely with the statistical values presented in table 2. 1 N 100.  k ,i .xm,i ACPxk ,i ,% = N  k (3) The variables selection (k) to find the best m =1  j k , j .xm, j QSPRMLR model depends upon the variation of the R2train, Q2CV and SE values.[1] A good model receives where N is the number of samples; m is the number of R2train and Q2CV meeting the statistical requirements (> samples for calculating Pxk,i values; k,i is the 0.6)[1] and should be close to 1.0, and the SE value regression coefficient of the equation. should be as small as possible. Table 2: The results of QSPRMLR models and the number variables (k) of 1 to 7 k Descriptors R²train R²adj Q²CV SE Fstat PRESS 1 x1 0.411 0.401 0.373 2.008 41.893 257.375 2 x1/x2 0.586 0.572 0.514 1.697 41.824 199.470 3 x1/x2/x3 0.709 0.694 0.638 1.435 47.171 148.580 4 x1/x2/x3/x4 0.755 0.738 0.680 1.328 43.960 131.477 5 x1/x2/x3/x4/x5 0.817 0.800 0.745 1.159 49.892 104.629 6 x1/x2/x3/x4/x5/x6 0.892 0.880 0.845 0.900 75.413 63.591 7 x1/x2/x3/x4/x5/x6/x7 0.901 0.889 0.833 0.866 70.532 68.523 Notation of molecular descriptors dipole x1 xc3 x5 5 C x2 ka1 x6 4 N x3 Total energy x7 fw x4 Contemporaneously, the variables of models because when the k value increase from 6 to 7, the must be statistically significant (p < 0.05). The data statistical values increase insignificantly; from table 2 and figure 2a indicated that when the k significantly, the Q2CV values decrease it means that values increase to 7, the MLR model takes the best the predictive ability of the model is reduced. So the statistical values; though, we get the k value of 6 MLR model is picked out with the statistical values © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 11
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Design, virtual screening and in silico… as equation (5). logβ12 = 19.893 + 0.417· x1 - 21.629 · x2 - 10.245 · x3 - 0.015 · x4 + 12.535 · x5 - 1.185 · x6 (5) n = 62; R2train = 0.892; Q2LOO = 0.845; SE = 9.000; Fstat = 75.413; PRESS = 63.591 1.0 20 2.00 0.9 1.75 0.8 16 R2train and Q2CV log12,exp 0.7 1.50 SE 0.6 12 1.25 0.5 1.00 0.4 8 0.75 0.3 0 1 2 3 4 5 6 7 8 8 12 16 20 (a) Number of variable, k (b) log 12,pred 2 2 Figure 2: (a) The changing of SE, R train and Q CV values according to k values; (b) Correlation of experimental (logβ12,exp) and predicted (logβ12,pred) values of the training data set using the QSPRMLR model with k = 6 Consequently, the training dataset of the built > ka1(x6) > xc3 (x5) are equal to the values of 39.305, QSPRMLR model is a perfect fit, and the correlation of 25.676 and 18.452, respectively. The fw parameter is experimental and predicted values of the training formula weight and is more significant; the molecular dataset using the QSPRMLR model with k of 6 is quite structure is cumbrous, which is the macromolecular fitting (figure 2b). This showed that the model effect. The ka1 index is the Kappa Alpha 1, which exhibits good predictive ability for the group of these explains the molecule presence of heteroatom (none- complexes. Thereupon, the model can be applied to carbon or none-hydrogen atoms).[15] This indicates it discover new complexes and develop ANN models quantifies the contribution degree of the heteroatom with the better predictive ability.[1] comparing a carbon sp3 atom. Meanwhile, the xc3 The study continued to use three models of the k index, namely Chi cluster 3, is the simple 3rd-order values of 5 to 7 to examine the contribution of cluster of chi indexes. According to graph theory,[1,18] variables of the built model via the GACPxi values, the order of a Chi index is the number of bonds in the which is the average value of ACPxk,i as equation (3). fragment of the molecule. These essential conclusions As results of Table 3 showed that the significant of the contribution of variables in the model allow the contribution of the descriptors in the order of fw (x4) development of new ligands and complexes. Table 3: The statistical values of QSPRMLR models and the evaluation of variables contribution in QSPRMLR models with k of 5 to 7 QSPRMLR ACPxk,i, % Statistical values GACPxi, % k=5 k=6 k=7 k=5 k=6 k=7 R2train 0.817 0.892 0.901 – – – – R2adj 0.800 0.880 0.889 – – – – Q2CV 0.745 0.845 0.833 – – – – SE 1.159 0.900 0.866 – – – – constant 16.905 19.893 18.762 – – – – x1 0.617 0.417 0.565 0.230 4.819 5.500 3.516 x2 -22.714 -21.629 -18.468 0.043 7.638 5.500 4.394 x3 -12.215 -10.245 -7.938 0.018 5.582 3.657 3.086 x4 -0.035 -0.015 -0.008 99.655 12.546 5.715 39.305 x5 6.720 12.535 8.917 0.055 34.614 20.687 18.452 x6 – -1.185 -1.712 – 34.802 42.226 25.676 x7 – – -0.003 – – 16.714 5.571 © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 12
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Tran Thai Hoa et al. 3.2. Constructing QSPRANN models (approximation of 15%).[16] The architecture of the neural network is I(6)-HL(m)-O(1), in which six In the first step, the QSPRANN models were developed neurons of the input layer I(6) are dipole, 5C, 4N, fw, on the six descriptors of the MLR model. The dataset, xc3, and ka1; the output layer with one neuron the including 62 experimental values, was randomly log12 values. The number of neurons of the hidden divided into a training subset of 44 values layer (m) was surveyed, and the initial values of the (approximation of 70%), a test subset of 9 values m neurons are indicated in table 4. (approximation of 15%), and a CV subset of 9 values Table 4: The initial built ANN model I(6)-HL(m)-O(1) with statistical criteria Training Test Validation Transfer Symbol QSPRANN R2train Q2test Q2CV error error error Function ANN1 I(6)-HL(6)-O(1) 0.903 0.921 0.980 0.669 0.566 0.251 tan-sigmoid ANN2 I(6)-HL(7)-O(1) 0.940 0.967 0.981 0.422 0.278 0.167 log-sigmoid ANN3 I(6)-HL(8)-O(1) 0.965 0.931 0.982 0.243 0.552 0.128 tan-sigmoid ANN4 I(6)-HL(3)-O(1) 0.958 0.966 0.980 0.297 0.261 0.184 tan-sigmoid ANN5 I(6)-HL(3)-O(1) 0.965 0.955 0.984 0.243 0.465 0.144 log-sigmoid 30 Q2 = 0.858 of QSPRMLR ext Q2 = 0.938 of QSPRANN4 ext 25 20 log12,exp 15 10 5 5 10 15 20 25 30 log12,pred (a) (b) Figure 3: (a) The description of QSPRANN I(6)-HL(3)-O(1) model; (b) Correlation of experimental versus predicted values on the external dataset of QSPR models In the second step, the QSPRANN models were independent data set.[1] As above-mentioned, this externally evaluated on a dataset of 20 observations study used the external data of 20 experimental to explore the premier QSPRANN model. The Q2ext and samples. The detailed data from 20 experimental MARE(%) are two indices to select the best model. observations and the calculated results are shown in So, the best QSPRANN model holds on suitable Q2ext table 5. values (> 0.5) [1], and the MARE(%) is the smallest As presented in table 5 and figure 4, the results value. Figure 4 indicates that three of five ANN indicated that the non-linear (QSPRANN4) and linear models, including ANN3, ANN4, and ANN5, satisfy (QSPRMLR) regression models confirmed the the Q2ext statistical values of 0.737, 0.938, and 0.644, correlation between the predicted and experimental respectively. In addition, the MARE value (%) of values with the Q2ext values of 0.938 and 0.858, these three ANN models is 15.124, 7,992, and 18.121, respectively. Accordingly, the predictive ability of respectively. Consequently, the ANN4 model the two models is the most approving.[1] received the lowest value of MARE (%), and the As another issue, the MARE (%) values of value of the Q2ext is close to 1.0. Thereupon, this ANN QSPRMLR and QSPRANN4 I(6)-HL(3)-O(1) models are model with the architecture of network I(6)-HL(3)- 11.681 and 7.992, respectively. The results indicate O(1) was selected to discover new complexes. that the QSPRANN models have better predictive ability than the QSPRMLR models and the predicted 3.3. External validation of models values (logβ12,cal) of QSPRANN model approximate the experimental values (logβ12,exp). The external validation (EV) technique is an The one-way ANOVA analysis is used to test the important stage in constructing a completely difference between the predicted values (logβ12,cal) predictive model, and it has to carry out on an and the beginning values (logβ12,exp) on the EV dataset © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 13
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Design, virtual screening and in silico… Table 5: The full information on external validation from the QSPR models Ligands Metal logβ12,cal logβ12,exp Ref. R1 R2 R3 R4 ions QSPRMLR QSPRANN4 H H -C5H4N -C5H4N Ni2+ 11.290 9.796 12.567 [19] H H -C5H4N -C5H4N Cu2+ 12.160 10.471 11.935 [19] H -CH3 -C5H4N -C5H4N Mn2+ 7.0000 7.364 7.265 [19] CH3 -CH3 -C5H4N -C5H4N Cu2+ 12.490 10.631 11.828 [19] H H -CH3 -C10H12NO Hg2+ 10.302 8.112 9.425 [20] H H -CH3 -C10H12NO Ag2+ 9.1984 7.919 8.209 [20] H H H -C6H3(OH)(OCH3) Cu2+ 18.050 18.520 17.413 [20] H H H - C6H4OH Hg2+ 23.820 18.539 20.065 [21] H H H -C10H6OH Mn2+ 9.9200 10.221 10.057 [21] H H H -C10H6OH Pb2+ 13.450 15.097 13.784 [21] H H H -C10H6OH Co2+ 13.650 11.834 11.886 [21] H H H -C10H6OH Ni2+ 14.370 12.335 12.860 [21] H H H -C10H6OH Cu2+ 15.310 14.969 15.850 [21] H H - -C9H7NO Ni2+ 15.876 19.986 17.918 [22] H H H -C6H4NH2 Ni2+ 23.700 22.747 22.216 [22] H H H -C6H4NH2 Co3+ 22.290 21.272 20.840 [23] H H H -C6H4NH2 Mn3+ 20.700 22.099 17.651 [23] H H H -C6H4NH2 Zn2+ 20.780 23.241 18.769 [23] H H H -C6H4NO2 Pr2+ 20.240 23.485 18.886 [23] H H H -C6H4NO2 Fe2+ 21.310 23.951 18.865 [23] MARE, % 11.681 7.992 140 MARE derivatives have been synthesized in practice and 1.0 Q2EV 0.938 have high biological activity, like thiosemicarbazone 120 0.858 0.8 and their complexes, to orient further future studies. 0.737 100 0.644 Six variables of the models, including dipole, 5C, 4N, MARE,% 80 0.560 0.6 fw, xc3, and ka1, were selected to discover the new Q2ext 0.526 thiosemicarbazone-based complexes. 60 0.4 In general, the two derivatives of phenothiazine 40 and carbazole were used to design new 25.704 0.2 20 11.681 20.819 15.124 18.121 thiosemicarbazone and their complexes among the 7.992 0 0.0 new ligands with some well-known metal ions such M LR N N 1 N N 2 N N 3 N N 4 N N 5 as Zn2+, Ni2+, Ag+, Cd2+, and Cu2+. The new ligands A A A A A QSPR models were set up by attaching the groups at the R4 site of Figure 4: The MARE (%) and Q2ext values of QSPR the ligand structure, although the rest sites, such as models R1, R2, and R3 of the initial structure, are hydrogen atoms. for two QSPRMLR and QSPRANN models. The results A variety of new thiosemicarbazone-based show that the difference between these two pairs of substances were designed and screened clearly by results is insignificant (F = 0.1591 < F0.05 = 3.1588). adding the same variables of new complexes into the initial training data set to find new complexes using 3.4. Development of new thiosemicarbazone-based AD and Outliers techniques through the D-Cook complexes absolute values (|D-Cook|).[1] As a result, 18 new thiosemicarbazone derivatives and 30 new ligands- This study looked for potential derivatives with based complexes were found, and calculated the similar properties to the original substance to develop stability constants (log12,new). The results are new ligands and complexes. Besides, these calculated on both of QSPRMLR and QSPRANN models © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 14
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Tran Thai Hoa et al. (table 6). Table 6. The calculated values (log12,new) from the constructed MLR and ANN models Metal logβ12,new Metal logβ12,new R4 site R4 site ions MLR ANN ions MLR ANN Zn2+ 13.956 13.939 Ni2+ 17.837 15.763 Ag+ 15.044 13.952 Cd2+ 7.743 7.986 Cd2+ 19.991 17.602 Cd2+ 16.655 15.229 16.599 Ag+ 17.6942 Ag+ 17.694 16.600 8 Cd2+ 11.190 9.816 Ag+ 8.030 7.924 2+ 2+ Ni 10.756 10.508 Cd 8.913 7.792 Ag+ 9.019 7.926 Cd2+ 8.451 7.581 Zn2+ 8.637 8.621 Zn2+ 8.640 7.668 Cu2+ 8.847 8.548 Ag+ 11.412 10.469 2+ 2+ Zn 9.316 8.559 Cd 11.142 9.780 Cu2+ 19.074 17.668 Cd2+ 19.239 16.801 Ni2+ 18.685 17.527 Ni2+ 17.606 16.932 Ag+ 17.262 17.826 Ag+ 19.953 18.385 Cu2+ 19.451 17.991 Cd2+ 19.406 18.075 Ni2+ 18.367 17.845 Zn2+ 20.871 18.376 Furthermore, the paired t-test analysis of two- 0.966, Q2test = 0.980 and Q2ext = 0.938). Also, the sample assuming unequal variances method was resulting models allowed the design of 18 new applied to evaluate the difference between the thiosemicarbazone ligands and 30 new metal- predictive results of the QSPRMLR and QSPRANN thiosemicarbazone complexes which are predicted models. It turns out that the difference is insignificant the stability constants (log12,new). The new substance (t-stat = 0.8639 < t-crit = 2.0009). can be used in required fields. In addition, results of QSPR-based models may be available for finding 4. CONCLUSION different thiosemicarbazone and their complexes. In this research, the QSPR modeling based on the Acknowledgment. Hue University supported this approaches as multivariate linear regression and work partially under the Core Research Program, artificial neural network is wholly developed through Grant No. NCM.DHH.2022.04. the data sets of the 62 stability constants values for the training set and the 20 stability constants values REFERENCES for the external set of thiosemicarbazone-based complexes. The novel PM7 and PM7/sparkle 1. R. Kunal, K. Supratik, N. D. Rudra A Primer on versions of semi-empirical quantum mechanics are QSAR/QSPR Modeling, Fundamental Concepts, used to optimize the structures of complexes. The best Springer, New York, USA, 2015. QSPRMLR model with six variables is constructed 2. T. Khan, R. Ahmad, S. Joshi, A. R. Khan. Anticancer with good statistical criteria (R2train = 0.892, Q2CV = potential of metal thiosemicarbazone complexes: a 0.845, SE = 0.900, and Q2ext = 0.858). In addition, the review, Chem. Sin., 6(12), 1-11, 2015. better ANN model with architecture I(6)-HL(3)-O(1) 3. S. Kumar, D. N. Dhar, P. N. Saxena. Applications of metal complexes of Schiff bases-A review, J. Sci. Ind. is developed from the variables of the QSPRMLR Res., 68(3), 181-87, 2009. model with excellent results (R2train = 0.958, Q2CV = © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 15
25728288, 2023, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/vjch.202200203 by Readcube (Labtiva Inc.), Wiley Online Library on [01/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Vietnam Journal of Chemistry Design, virtual screening and in silico… 4. N. M. Quang, T. X. Mau, N. T. A. Nhung, T. N. M. acetylpyridinethiosemicarbazone (2-APT), An, P. V. Tat. Novel QSPR modeling of stability Polyhedron., 7(2), 147-150, 1998. constants of metal-thiosemicarbazone complexes by 14. Sahadev, R. K. Sharma, S. K. Sindhwani. hybrid multivariate technique: GA-MLR, GA-SVR Potentiometric Studies on the Complexation and GA-ANN, J. Mol. Struct., 1195, 95-109, 2019. Equilibria Between Some Trivalent Lanthanide Metal 5. M. H. Keshavarz, Z. Shirazi, A. Barghahi et al. A Ions and Biologically Active 2-Hydroxy-1- novel model for prediction of stability constants of the Naphthaldehyde Thiosemicarbazone (HNATS), thiosemicarbazone ligands with different types of Monatshefte fur Chemie., 123, 883-889, 1992. toxic heavy metal ions using structural parameters and 15. B.S. Garg, S. Ghosh, V. K. Jain, P. K. Singh. multivariate linear regression method., Environ Sci Evaluation of thermodynamic parameters of bivalent Pollut Res., 29, 37084-37095, 2022. metal complexes of 2-hydroxyacetophenone- 6. N. M. Quang, N. T. A. Nhung, P. V. Tat. An insight thiosemicarbazone (2-HATS), Thermochimica Acta., QSPR-based prediction model for stability constants 157, 365-368, 1990. of metal-thiosemicarbazone complexes using MLR 16. Mohammad Najjarpour, Hossein Jalalifar, Saeid and ANN methods, Vietnam J. Chem., 57, 500-506, Norouzi-Apourvari. Half a century experience in rate 2019. of penetration management: Application of machine 7. S. V. Babu, K. H. Reddy. Direct Spectrophotometric learning methods and optimization algorithms - A Determination of Mercury (II) using 2-Acetylpyridine review, Journal of Petroleum Science and Thiosemicarbazone in Environmental Samples, Indian Engineering, 208, Part D, 109575, 2022. J. Adv. Chem. Sci., 1, 65-72, 2012. 17. J. Gasteiger, J. Zupan. Neural Networks in Chemistry, 8. J. J. P. Stewart. Optimization of parameters for Chiw. Inr. Ed. EngI., 32, 503-521, 1993. semiempirical methods VI: more modifications to the 18. L. H. Hall, L. B. Kier. The Molecular Connectivity Chi NDDO approximations and re-optimization of Indices and Kappa Shape Indices in Structure- parameters, J. Mol. Model., 19, 1-32, 2013. Property Relations, in Reviews of Computational 9. D. D. Steppan, J. Werner, and P. R. Yeater. Essential Chemistry, Donald Boyd and Ken Lipkowitz, Eds. Regression and Experimental Design for Chemists Location: VCH Publishers, 1991. and Engineers. Free Software Package, 1998. 19. P. V. Bernhardt, P. C. Sharpe, M. Islam, D. B. 10. D. N. Reddy, K. V. Reddy, B. M. Tegegne, V. K. Lovejoy, D. S. Kalinowski, D. R. Richardson. Iron Reddy. Development of a Highly Sensitive Extractive Chelators of the Dipyridylketone Thiosemicarbazone Spectrophotometric Method for the Determination of Class: Precomplexation and Transmetalation Effects Nickel(II) from Environmental Matrices Using 2- on Anticancer Activity, J. Med. Chem., 52(2), 407- Acetylpyridine-4-methyl-3-thiosemicarbazone, 415, 2009. American J. Anal. Chem., 3, 719-726, 2012. 20. Atalay, Ozkan. Thermodynamic studies of some 11. K. Syamasundar, M. A. Chari, D. Shobha. Extractive complexes of 4’-morpholino acetophenone Spectrophotometric Determination of Cobalt(II) With thiosemicarbazone, Thermochimica Acta, 237, 369- 1-(Benzimidazol-2-Yl) Ethanone Thiosemicarbazone, 374, 1994. Anal. Chem. Indian J., 2(2), 45-47, 2006. 21. Sahadev, R. K. Sharma, S. K. Sindhwani. Thermal 12. R. H. U. Borges, E. Paniago, H. Beraldo. Equilibrium studies on the chelation behaviour of biologically and Kinetic Studies of Iron(II) and Iron(III) active 2-hydroxy-1-naphthaldehyde thiosemi- Complexes of Some a(N)-Heterocyclic carbazone (HNATS) towards bivalent metal ions: a Thiosemicarbazones. Reduction of the Iron(III) potentiometric study, Thermochimica Acta, 202, 291- Complexes of 2-Formylpyridine Thiosemicarbazone 299, 1992. and 2-Acetylpyridine Thiosemicarbazone by Cellular 22. K. Sarkar, B. S. Garg. Determination of Thiol-Like Reducing Agents, J. Inor. Biochem., 65, thermodynamic parameters and stability constants of 267-275, 1997. the complexes of p-MITSC with transition metal ions, 13. B. S. Garg, S. R. Singh, S. B. Basnet, R. P. Singh. Thermochimica Acta, 113, 7-14, 1987. Potentiometric studies on the complexation equilibria 23. S. S. Sawhney, and S. K. Chandel. Solution chemistry between La(III), Pr(III), Nd(III), Gd(III), Sm(III), of Cu(II)-, Co(II)-, Ni(II)-, Mn(II)- and Zn(II)-p- Tb(III), Dy(III), Ho(III) and 2- aminobenzaldehyde thiosemicarbazone systems, Thermochimica Acta, 71, 209-214, 1983. Corresponding authors: Tran Thai Hoa Pham Van Tat Faculty of Chemistry, Hue University of Sciences Department of Sciences and Journal Management Hue University, 77 Nguyen Hue, Hue City, 49100, Hoa Sen University, 8 Nguyen Van Trang Viet Nam Dist. 1, Ho Chi Minh City 700000, Viet Nam Email: tthaihoa@hueuni.edu.vn. E-mail: vantat@gmail.com. © 2023 Vietnam Academy of Science and Technology, Hanoi & Wiley-VCH GmbH www.vjc.wiley-vch.de 16