Predicting estrogen activities of bisphenol a and its analogs using quantum chemistry calculations and artificial neural networks Текст научной статьи по специальности «Медицинские технологии»

УДК: 543.683

ПРОГНОЗИРОВАНИЕ ЭСТРОГЕННОЙ АКТИВНОСТИ БИСФЕНОЛА А И ЕГО АНАЛОГОВ

С ИСПОЛЬЗОВАНИЕМ КВАНТОВО-ХИМИЧЕСКИХ ВЫЧИСЛЕНИЙ И ИСКУССТВЕННЫХ

НЕЙРОННЫХ СЕТЕЙ

Ву Ван Дат, Ле Ким Лонг, Доан Ван Фук, Нгуен Хоанг Чанг, Нгуен Ван Чанг, Нгуен Тхи Тху Ха

Ву Ван Дат *

Химический Факультет, Институт Eстественных Наук, Вьетнамский Национальный Университет, Ханой, 19 Ле Тхань Тонг, Хоан Кием, Ханой, Вьетнам Email: vvdat@most.gov.vn*

Ле Ким Лонг, Нгуен Хоанг Чанг

Педагогический Факультет, Институт Образования, Вьетнамский Национальный Университет, Ханой, 144 Суан Тху, Кау Зьау, Ханой, Вьетнам Email: longlk@vnu.edu.vn, hoangtrang@mail.ru

Доан Ван Фук

Институт химии и материалов, 17 Хоанг Шам, Кау Зьау, Ханой, Вьетнам. Email: doanphucspbstu@mail.ru

Нгуен Ван Чанг

Институт тропической технологии, Вьетнамская академия наук и технологий, 18 Хоанг Суок Вьет, Кау

Зьау, Ханой, Вьетнам

Email: nguyenvantrangsphn@gmail.com

Нгуен Тхи Тху Ха

Ханойский национальный университет образования, 136 Суан Тху, Кау Зьау, Ханой, Вьетнам. Email: ntt.ha@hnue.edu.vn

В данной работе представлены результаты исследования количественной взаимосвязи между структурой и активностью (QSAR) бисфенола А и его аналогов с использованием квантовых химических расчетов и метода искусственных нейронных сетей (ANN). Анализ молекулярной структуры был выполнен на основе теории функционала плотности (DFT) методом B3LYP/6-31+G(d). Квантовые химические расчеты концентрируются на оптимизировании молекулярной структуры, частоты колебаний и энергии молекулярной орбиты. Распределение электронной плотности атомов было изучено в рамках естественного орбитального анализа связи (NBO). Все расчеты были выполнены в программных пакетах Gaussian 9 и Gaussview 6. Полученные в результате расчетов структурно-квантовые параметры и ранее известные экспериментально наблюдаемые биологические данные об эстро-генной активности рабочей выборки были использованы в качестве входных данных для построения модели QSAR с использованием метода искусственных нейронных сетей. С помощью метода искусственных нейронных сетей были выявлены также структурно-квантовые параметры, наиболее эффективно влияющие на эстрогенную активность исследованных веществ. Для проверки эффективности и стабильности построенных моделей был использован метод скользящего контроля и формирования внешней тестовой выборки. Полу-ченые QSAR модели имеют довольно хорошие статистические параметры: R2 = 0,99; Q2loo = 0,98; R2Predict = 0,98. В соответствии с этим результатом следует отметить, что предлагаемая нами модель, построенная методом искусственных нейронных сетей с использованием параметров квантовой химии, достаточно адекватна и может быть полезна для прогнозирования эстрогенной активности неизученных производных и аналогов BPA с достаточной надежностью.

Ключевые слова: количественные соотношения «структура-свойство», модель QSAR, бисфенол А, эстрогенная активность, теория функционала плотности, 6-31+G(d) базисная функция, метод искусственных нейронных сетей

PREDICTING ESTROGEN ACTIVITIES OF BISPHENOL A AND ITS ANALOGS USING QUANTUM CHEMISTRY CALCULATIONS AND ARTIFICIAL NEURAL NETWORKS

Vu Van Dat, Le Kim Long, Doan Van Phuc, Nguyen Hoang Trang, Nguyen Van Trang, Nguyen Thi Thu Ha

Vu Van Dat*

Faculty of Chemistry, University of Science, Vietnam National University, Le Thanh Tong st., 19, Hoan Kiem Dist., Hanoi, Vietnam Email: vvdat@most.gov.vn*

Le Kim Long, Nguyen Hoang Trang

Faculty of Education, University of Education, Vietnam National University, Xuan Thuy st., 144, Cau Giay Dist., Hanoi, Vietnam

Email: longlk@vnu.edu.vn, hoangtrang@mail.ru Doan Van Phuc

Institute of Chemistry and Material, Hoang Sam st., 17, Cay Giay Dist., Hanoi, Vietnam Email: doanphucspbstu@mail.ru

Nguyen Van Trang

Institute for Tropical Technology, Vietnam Academy of Science and Technology, Hoang Quoc Viet st., 18, Cau

Giay Dist., Hanoi, Vietnam

Email: nguyenvantrangsphn@gmail .com

Nguyen Thi Thu Ha

Hanoi National University of Education, 136 Xuan Thuy st., Cau Giay, Hanoi, Vietnam Email: ntt.ha@hnue.edu.vn

This article presents the results of the quantitative structure - activity relationship (QSAR) study of bisphenol A (BPA) and its analogs using quantum chemistry calculations and method of artificial neural networks (ANN). Molecular structural analysis is performed using Density Functional Theory (DFT) at the B3LYP/6-31+G(d) level. The quantum calculations focus on finding the optimized molecular structures, vibrational frequencies, the molecular orbital energies with reasonable accuracy. The study of electron density distribution was carried out in the framework of the natural bond orbital (NBO) methods. The obtained parameters and known observable estrogen activities are used as input data for constructing the QSAR model, using the artificial neural network method. Based on the artificial neural network method the quantum parameters having the strongest impact on the estrogen activity of the compounds were revealed. The internal and external validation methods have been performed to test the performance and the stability of the model. The statistical parameters obtained of the QSAR model were: R2 = 0.99; Q2loo = 0.98; R2Predict = 0.98. According to the obtained results, our proposed model, constructing by method of artificial neural network using the parameters of quantum chemistry is adequate and may be useful to predict of estrogen activities for unexplored derivatives and BPA analogs with m od-erate reliability.

Key words: quantitative structure - activity relationship, QSAR model, bisphenol A, estrogen, Density Functional Theory, 6-31+G(d) basis function, artificial neural networks - ANN

Для цитирования:

Ву Ван Дат, Ле Ким Лонг, Доан Ван Фук, Нгуен Хоанг Чанг, Нгуен Ван Чанг, Нгуен Тхи Тху Ха Прогнозирование эстрогенной активности бисфенола А и его аналогов с использованием квантово -химических вычислений и искусственных нейронных сетей. Изв. вузов. Химия и хим. технология. 2019. Т. 62. Вып. 5. С. 31-37 For citation:

Vu Van Dat, Le Kim Long, Doan Van Phuc, Nguyen Hoang Trang, Nguyen Van Trang, Nguyen Thi Thu Ha Predicting estrogen activities of bisphenol A and its analogs using quantum chemistry calculations and artificial neural networks. Izv. Vyssh. Uchebn. Zaved. Khim. Khim. Tekhnol. 2019. V. 62. N 5. P. 31-37

INTRODUCTION

Quantitative structure - activity relationship (QSAR) is a statistical research method that establishes empirical models describing relationship between the biological activity of compounds and their chemical structures. In fact, the structure-activity relationship is often complex and difficult to express by mathematical functions [1]. The application of classical data processing methods (Multivariate Regression Analysis, Partial Least Squares, etc.) to QSAR models is often problematic. The requirement for QSAR researchers is to apply state-of-the-art data processing and statistical analysis methods through specialized software to create a highly stable and predictable model. One of the most widely used data processing methods today is artificial neural network (ANN).

In this article, we developed a QSAR model for Bisphenol A and its analogs using quantum chemistry and artificial neural networks. The basis for selecting this substance for investigation is derived from the need to use BPA and its analogs [2, 3] as well as their severity [4-9].

METHODS AND MATERIALS

A general schema of QSAR method is showed in Fig. 1.

Fig. 1. Schema of QSAR method Рис. 1. Схема построения QSAR модели

2.1. Biological data

The data set used in this study include 23 compounds synthesized and their biological activities analyzed by a team of researchers from the University of Minnesota and the University of New Orleans, USA [10]. The biological activity selected for this study is estrogen level of activity evaluated in the form of biological expression of the reporter genes embedded in the cell. The structures of the molecules in the data set are presented in Table 1.

Table 1

Structures of the molecules in the data set [10] Таблица 1. Структуры изучаемых молекул, входящих в набор данных_

Compound

Substituents

1 2 3 4 5 6 7 8 9 R1 R2

1 DM DMB Bis A H CH3 OH H H H CH3 OH H CH3 CH2CH(CH3)2

2 DMB Bis A H H OH H H H H OH H CH3 CH2CH(CH3)2

3 MM4 H H OH H H H H OH H C2H5 C2H5

4 Bis A H H OH H H H H OH H CH3 CH3

5 HF Bis A H H OH H H H H OH H CF3 CF3

6 DM HPTE H CH3 OH H H H CH3 OH H H CCl3

7 MM1 H H OH H H H H OH H H CH3

8 Bis F H H OH H H H H OH H H H

9 Bis B H H OH H H H H OH H CH3 C2H5

10 DM Bis A H CH3 OH H H H H OH CH3 CH3 CH3

11 HPTE H H OH H H H H OH H H CCl3

12 1844-00-44 H H OH H H H H OH H H CH(CH3)2

13 MM2 H H OH H H H H OH H H C2H5

14 TM Bis A H CH3 OH CH3 H H CH3 OH CH3 CH3 CH3

15 o,p'-Bis A H H H H OH H H OH H CH3 CH3

Compound 1 2 3 4 5 6 7 8 9 R1 R2

16 Mono Mxy Bis A H H OH H H H H OCH3 H CH3 CH3

17 P Bis A H H OH H H H H OH H CH3 C6H5

18 PCP H H H H H H H OH H CH3 CH3

19 MH MM1 H H OH H H H H H H CH3 H

20 MH Bis F H H H H H H H OH H H H

21 TC Bis A H Cl OH Cl H H Cl OH Cl CH3 CH3

22 TB Bis A H Br OH Br H H Br OH Br CH3 CH3

23 Mxy Bis A H H OCH3 H H H H OCH3 H CH3 CH3

2.2. Structural parameter calculation The quantum parameters characterizing the molecular structure are calculated based on the Density Functional Theory (DFT) at B3LYP/6-31+G(d) level [11-13]. The calculation methods are implemented in the Gaussian 09 software [12]. The visualization of the studied molecules and the initial structures are conducted using ChemCraft software [14]. The list of structural quantum parameters calculations is presented in Table 2.

Table 2

List of quantum descriptorsof of bisphenol A analogs Таблица 2. Список квантовых дескрипторов исследованных веществ

2.3. Statistical analysis and variable selection In this article, three-layer artificial neural network (input layer, hidden layer and output layer) with feedforward technique is used. Accordingly, each node of the preceding layer links with all nodes in the following layers and nodes in the same layer are not linked. The number of nodes entered by independent variable numbers will be used; the number of nodes of the output layeris a representative value of the activity

of the molecule being investigated; the number of nodes in the hidden layer is set by the network regulator in such a way that all information in the data set is retrieved, but no over-processing occurs. This is also the most common type of ANN used in the QSAR model [15-18].

The operation diagram of ANN is illustrated in Fig. 2, where p is the input data vector; w is the weight typical for the link between data (signal) transmitted and data (signal) received; f is the transfer function. In the ANN network, the function f is usually a hard-limit, linear, log-sigmoid or tan-sigmoid function.

Hidden Layer Output Layer

ID 1

Fig. 2. Diagram of artificial neuron network Рис. 2. Схематическая иллюстрация искусственных нейтронных сеток

However, not all of the calculated structural parameters are statistically significant for the model, therefore,selective evaluation of the potential variables to construct the structural data sets must be conducted. The process of selecting variables was performed by examining the correlation between structural parameters through the Pearson correlation matrix to find parameters that are closely related. Next, the set of structural parameters continues to be filtered through sensitivity about the mean procedure with the activity value. The nature of this operation is to investigate the variability of the activity according to the variability of an independent variable while keeping the values of other independent variables. These procedures are implemented through the NeuroSolution 6.0 program and the NeuroSolution for Excel add-in [11].

Testing the stability, predictability and generality of models a crucial step in QSAR research. The quality and predictability of the models is assessed by internal and external validation through statistical indicators [19, 20]. Cross-validation is a common method of referencing. Accordingly, 23 compounds were ran-

Symbol Definition Unit

Length Bond length Â

Angel Bond angle Degree

Dihedral Dihedral Degree

¿HOMO Energy of the Highest Occupied Molecular Orbital eV

-Èlumo Energy of the lowest unoccupied Molecular Orbital eV

И Dipole Moment Debye

Esp Total Energy of Molecular eV

AE AE = Elumo - Ehomo eV

X „, _ _((lumo + Ehomo ) - absolute Z 2 electronegativity eV

n elumo ~ehomo - absolute hard-2 ness eV

œ m_f1/ - reactivity index / 2^ eV

C1, C2, ... C13 Atomic charges on carbon atoms 1,2,.. respectively e

domly divided into 3 sets: training set (70% of molecules); validation set (15% of molecules) and test set (15% of elements). The predictability of the model is evaluated through the leave-three-out procedure with the characteristic parameter as the generalization coefficient Q2. The more Q2 approaches 1, the more stable and predictable the model is. However, according to Tropsha and other experts in the field of QSAR research, internal modeling indicators do not guarantee the predictability for new compounds outside the model [19]. A quality and predictable QSAR model must be built with external references. The model's external references is made by using the model constructed from training set to predict the activity of the compounds in the test set, and then the predicted values are computed to calculate out-of-model indicators.

For ANN, the cessation criterion is that the Mean Square Error (MSE) of the validation indicates an increase; training algorithm is Levenberg-Marquardt; number of neurons on hidden layers is examined from 1 to 10; MSE values from 10 ANN models will be compared to select the optimum network for the smallest MSE.

RESULTS AND DISCUSSION

We calculated 50 characteristic parameters of 23 molecules. These parameters were compared with

the corresponding parameters of BPA molecule calculated using the same method. The results indicate that there are 18 parameters that have zero descriptor. The remaining parameters were selected through the Pearson correlation coefficient matrix and their sensitivity about the mean values with activity values.

Fig. 3. Schema illustrating the sensitivity of parameters to activity Рис. 3. Чувствительность дескрипторов на биоактивность

Based on the Pearson correlation matrix and the schema showing the sensitivity of the parameters with activity (Fig. 3), the selection of the 10 most sensitive parameters could be considered independent variables set up to build the model. Data on these parameters are given in Table 3.

Table 3

Таблица.

Structural parameters selected to build the model

Compound C11 Ehomo C3 Ц C13 C6 C12 P C5 Esp

DMB Bis A -0.048 -5.92166 0.303 1.625279 -0.049 -0.209 -0.1 0.36884 -0.205 -849.63207

HPTE -0.069 -6.23404 0.315 1.877432 -0.079 -0.201 -0.304 0.41546 -0.194 -2071.14844

MM4 -0.052 -5.98588 0.303 2.2785 -0.052 -0.218 -0.098 0.36794 -0.218 -810.31824

DM DMB Bis A -0.04 -5.79295 0.308 0.922172 -0.04 -0.203 -0.09 0.36609 -0.206 -928.27007

HF Bis A -0.095 -6.55186 0.319 1.5767 -0.095 -0.198 -0.032 0.35784 -0.198 -1327.16975

Bis B -0.055 -5.96601 0.303 2.076508 -0.052 -0.214 -0.105 0.36975 -0.212 -771.00760

DM Bis A -0.045 -5.77608 0.308 1.9216 -0.045 -0.215 -0.112 0.37007 -0.215 -810.33368

P Bis A -0.067 -5.82533 0.349 1.30318 -0.066 -0.207 -0.077 0.37783 -0.203 -923.46071

MM2* -0.063 -5.95268 0.303 2.048301 -0.061 -0.215 -0.282 0.37456 -0.21 -731.69810

Bis A -0.052 -5.93281 0.302 1.7296 -0.052 -0.208 -0.113 0.36948 -0.208 -731.69781

PCP -0.052 -6.06833 -0.248 1.545729 -0.021 -0.208 -0.115 0.35871 -0.228 -656.47670

TM Bis A -0.037 -5.66941 0.312 1.3057 -0.037 -0.212 -0.111 0.36796 -0.212 -888.97147

MH MM1 -0.057 -6.07839 -0.247 1.535246 -0.028 -0.21 -0.29 0.35936 -0.225 -617.16711

o.p'-Bis A -0.039 -5.90125 -0.228 1.766342 -0.068 -0.213 -0.111 0.36733 0.318 -731.69522

MH Bis F -0.069 -6.11730 -0.247 1.445198 -0.039 -0.208 -0.479 0.36065 -0.229 -577.85490

MM1 -0.057 -5.95105 0.303 1.439828 -0.059 -0.211 -0.287 0.36842 -0.205 -692.38812

Bis F -0.069 -5.98996 0.302 1.4703 -0.069 -0.209 -0.476 0.37013 -0.209 -653.07586

DM HPTE -0.062 -6.07676 0.321 2.302561 -0.073 -0.197 -0.303 0.42385 -0.2 -2149.78590

1844-00-44* -0.051 -5.92819 0.302 0.632798 -0.051 -0.212 -0.269 0.37605 -0.21 -771.01496

Mono MxyBis A -0.066 -5.67513 0.335 0.943963 -0.063 -0.199 -0.07 0.36812 -0.203 -771.01928

TC Bis A -0.03 -6.55785 0.283 3.5848 -0.03 -0.224 -0.1 0.37544 -0.224 -2570.07512

TB Bis A -0.029 -6.51676 0.281 3.5184 -0.029 -0.221 -0.1 0.38069 -0.221 -11016.21701

MxyBis A* -0.051 -5.79866 0.306 0.9166 -0.051 -0.212 -0.113 0.37028 -0.212 -810.30816

The network training start with the 10 independent variables identified above. The MSE values of the ANN models with the number of neurons on the hidden layer varies from 1 to 10 are analyzed. Compare the MSE values of the models, we could see that the ANN network with 9 neurons on the hidden layer provide the most appropriate MSE value for training set, testset and validation set (R2 = 0.99, Q2 = 0.98, Rtes? = 0.98) (Fig. 4).

The correlation between the predicted, estimated and observed values are demonstrated on the graphs in Fig. 5.

Processing elements in hidden I

Fig. 4. Variabilities of MSE by number of neurons on the hidden layer Рис. 4. Валидация MSE от числа элементов скрытого слоя

The calculated and predicted activity values for the ANN are given in Table 4.

Table 4

Data on the biological activity of the substances Таблица 4. Данные о биологической активности исследованных веществ

Compound LgECsa (Observed) [10] LgEC5o(Predicted)

DMB Bis A -2.03 -2.05

HPTE -3.37 -3.35

MM4 -2.28 -2.30

DM DMB Bis A -1.99 -2.07

HF Bis A -2.79 -2.77

Bis B -3.28 -3.30

DM Bis A -3.31 -3.29

P Bis A -4.05 -4.03

MM2* -3.57 -3.76

Bis A -2.56 -2.57

PCP -4.05 -4.04

TM Bis A -3.8 -3.99

MH MM1 -4.05 -4.03

o.p'-Bis A -3.96 -3.93

MH Bis F -4.05 -4.04

MM1 -3.15 -3.16

Bis F -3.28 -2.91

DM HPTE -2.91 -2.50

1844-00-44* -3.38 -3.48

Mono Mxy Bis A -4.04 -3.90

TC Bis A -6.04 -5.94

TB Bis A -6.04 -5.96

Mxy Bis A* -6.04 -5.96

* - Molecule in test set

* - Молекула в тестовом наборе

b

Fig. 5. Correlation between predicted and observed values: Рис. 5. Сопоставление прогнозированных значений биоактивности с экспериментальными данными:

The QSAR model built by ANN method has a good predictability with R2test = 0.98. This model not only allows the establishment of a quantitative relationship between the structure and estrogen activity of molecules, but also helps to identify the parameters that have a great effect on the activity. According to the calculation results these parameters are C11, Ehomo, C3, ц, C13, C12 and C6. Furthermore, a structural change often entail a rigorous and general change of many quantum parameters. Therefore, the increase or decrease of a particular parameter in the structures of BPA and its derivatives does not reflect consistently in the change of estrogen activity. Based on the established QSAR models, it is necessary to analyze the change of the parameter of several molecules with high structural similarities. Thereby, it is possible to select optimal substituents on the molecular frame to establish new molecules with better biological response.

CONCLUSION

The QSAR study of BPA and its derivatives in this article is done through quantum chemistry calculations using B3LYP/6-31+G(d) method combined with modern data processing (ANN method). The survey results obtained a sustainable QSAR model with a determination coefficient R2 = 0.99. The generalization and exogenous capabilities of this model are at a high level with a generalization coefficient Q2 = 0.98; R2test =

0.98. so that this model can be applied in practice to predict the activity of BPA derivatives that haven't been studied.

REFERENCES ЛИТЕРАТУРА

1. Jalali-Heravi M. Neural Networks in Analytical Chemistry. In: D. Livingstone (Ed.) Artificial Neural Networks. Humana Press. 2009. P. 78-118. DOI: 10.1007/978-1-60327-101-1_6.

2. Morgan M.K., Jones P.A., Calafat A.M., Ye X., Croghan C.W., Chuang J.C., Wilson N.K., Clifton M.S., Figueroa Z., Sheldon L.S. Environ. Sci. Technol. 2011. 45. P. 5309. DOI: 10.1021/es200537u.

3. Schecter A., Malik N., Haffner D., Smith S., Harris T.R., Paepke O., Birnbaum L. Environ. Sci. Technol. 2010. 44. P. 9425. DOI: 10.1021/es102785d.

4. Vandenberg L.N., Hauser R., Marcus M., Olea N., Welshons W.V. Reproduct. Toxicol. 2007. 24. P. 139-177. DOI: 10.1016/j.reprotox.2007.07.010.

5. Krishnan A.V., Stathis P., Permuth S.F., Tokes L., Feld-man D. Endocrine Soc. 1993. V. 132. N 6. DOI: 10.1289/ehp.95103s7129.

6. Rezg R, El-Fazaa S, Gharbi N, Mornagui B. Bisphenol A and human chronic diseases: Current evidences, possible mechanisms, and future perspectives. Envir. Internal 2014. 64. P. 83-90. DOI: 10.1016/j.envint.2013.12.007.

7. Doerge D.R., Twaddle N.C., Vanlandingham M., Brown R.P., Fisher J.W. Toxicol. Appl. Pharmacol. 2011. 255. P. 261. DOI: 10.1016/j.taap.2011.07.009.

8. Ho S.M., Tang W.Y., Belmonte de Frausto J., Prins G.S. Developmental exposure to estradiol and bisphenol A increases susceptibility to prostate carcinogenesis and epigenet-ically regulates phosphodiesterase type 4 variant 4. Cancer

Res. 2006. 66 (11). P. 5624-32. DOI: 10.1158/0008-5472.CAN-06-0516.

9. Rochester J.R., Bolden A.L. Bisphenol S and F: A Systematic Review and Comparison of the Hormonal Activity of Bisphenol A Substitutes. Environ. Health Perspect. 2015. 123(7). P. 643-650. DOI: 10.1289/ehp.1408989.

10. Kelly P.C., William A.T., Thomas E.W. QSAR models of the in vitro estrogen activity of bisphenol A analogs. QSAR Comb.Sci. 2003. 22. P. 78—88. DOI: 10.1007/s11434-006-0287-1.

11. Zhao Y., Truhlar D. The M06 suite of density functionals for main group thermochemistry, thermochemical kinetics, non-covalent interactions, excited states, and transition elements: two new functionals and systematic testing of four M06-class functionals and 12 other functionals. Theor. Chem. Account. 2008. 120 (1-3). P. 215-241. DOI: 10.1007/s00214-007-0310-x.

12. Frisch M.J.T. Gaussian 09. Revision D.01. Gaussian, Inc. Wallingford CT. 2009.

13. Hmamouchi R., Larif M., Adad A., Bouachrine M., Lakhlifi T. QSAR analysis of the toxicity of phenols and thi-ophenols using MLR and ANNJ. Comp. Meth. Mol. Des. 2014. 4(3). P. 61-71. DOI: 10.1016/j.jtusci.2016.03.002.

14. Andrienko G.A. Chemcraft. https://www.chemcraftprog.com.

15. Devillers J. Strengths and Weaknesses of the Backpropagation Neural Network in QSAR and QSPR Studies. In: J. Devillers (Ed.) Neural Networks in QSAR and Drug Design. London: Academic Press. 1996. P. 1-46. DOI: 10.1016/B978-012213815-7/50002-9.

16. Hornik K. AppSroximation capabilities of multilayer feedforward networks. Neural Networks. 1991. 4. P. 251-257. DOI: 10.1016/0893-6080(91 )90009-T.

17. Balabin R.M., Safieva R.Z., Lomakina E.I. Comparison of linear and nonlinear calibration models based on near infrared (NIR) spectroscopy data for gasoline properties prediction. Chemometr. Intell. Lab. 2007. 88 (2). P. 183-188. DOI: 10.1016/j. chemolab.2007.04.006.

18. Tahmasebi P., Hezarkhani A. Application of a Modular Feedforward Neural Network for Grade Estimation. Nat. Resour. Res. 2011. 20. (1). P. 25-32. DOI: 10.1007/s11053-011-9135-3.

19. Golbraikh A., Tropsha A. Beware of q2. J. Mol. Graphics Model. 2002. 20. P. 269-276. DOI: 10.1023/A:1025386326946.

20. Steyerberg E.W., Bleeker S.E., Moll H.A., Grobbee D.E., Moons K.G. Internal and external validation of predictive models: A simulation study of bias and precision in small samples. J. Clinic. Epidemiol. 2003. V. 56. N 5. P. 441-447. DOI: 10.1016/S0895-4356(03)00047-7.

Поступила в редакцию 06.12.2018 Принята к опубликованию 07.03.2019

Received 06.12.2018 Accepted 07.03.2019