Application Note pubs.acs.org/jcim
MetaTox: Web Application for Predicting Structure and Toxicity of Xenobiotics’ Metabolites Anastasia V. Rudik,*,† Vladislav M. Bezhentsev,† Alexander V. Dmitriev,† Dmitry S. Druzhilovskiy,† Alexey A. Lagunin,†,‡ Dmitry A. Filimonov,† and Vladimir V. Poroikov† †
Institute of Biomedical Chemistry (IBMC), 10 Bldg. 8, Pogodinskaya Str., Moscow, 119121, Russia Medico-biological Faculty, Pirogov Russian National Research Medical University, 1 Ostrovityanova Str., Moscow, 117997, Russia
‡
ABSTRACT: A new freely available web-application MetaTox (http:// www.way2drug.com/mg) for prediction of xenobiotic’s metabolism and calculation toxicity of metabolites based on the structural formula of chemicals has been developed. MetaTox predicts metabolites, which are formed by nine classes of reactions (aliphatic and aromatic hydroxylation, Nand O-glucuronidation, N-, S- and C-oxidation, and N- and O-dealkylation). The calculation of probability for generated metabolites is based on analyses of “structure-biotransformation reactions” and “structure-modified atoms” relationships using a Bayesian approach. Prediction of LD50 values is performed by GUSAR software for the parent compound and each of the generated metabolites using quantitative structure−activity relationahip (QSAR) models created for acute rat toxicity with the intravenous type of administration.
■
INTRODUCTION In the human organism, many xenobiotics represented by small organic molecules are modified during biotransformation, which can be classified into three phases. Usually, phase I activates the chemical by introducing a reactive and polar functional group, whereas phase II conjugates the activated compound with a charged species, increasing the molecular weight, reducing reactivity, and improving the transport property. During phase III transformed xenobiotics are eliminated from the cell into the extracellular medium. Hydroxylation and dealkylation are the most common phase I reactions. The majority of phase I reactions (about two-thirds in the case of approved drugs) are catalyzed by cytochromes P450.1,2 Cytochromes P450 play important roles in detoxification and deactivation of drugs.3 However, P450 also can transform the parent compound into the more toxic metabolite.4,5 Phase II is presented by conjugation reactions. Glucuronidation is the most common phase II reaction. It is catalyzed by UDP-glucuronosyltransferases. More polar metabolites are formed as a result of phase II reactions. These metabolites are actively transported across membranes. Metabolites, which are formed during biotransformation, can be significantly different from the parent compound on bioactivity profiles, cause enhanced efficacy or a loss of activity, induce drug−drug interactions, or adverse and toxic effects.1 Therefore, it is important to evaluate the structures of metabolites at early stages of drug development. Prediction of the structure of metabolites can be performed in different ways. The first way is to predict the sites of metabolism (SOMs)the atoms or groups of atoms that are attacked by enzymes and transformed during biotransforma© 2017 American Chemical Society
tion. This way is focused on the different enzymes, and knowledge of the enzyme mechanism of action due to the results of SOMs prediction provides hypotheses regarding the possible structures of metabolites. There are many computational tools that can predict SOMs for the different enzymes, mostly for cytochromes P450, 6−8 but also for UDPglucuronosyltransferases.8,9 The second way is to predict the structure of metabolite without preliminary prediction of SOMs. Primarily it is realized in expert systems, which predict products of metabolism using dictionaries of biotransformation.10 Such systems contain rules of transformation of parent compounds into their metabolites. The principal drawback of this approach is a generation of a combinatorial explosion of predictions, as all possible combinations of transformations permitted by the dictionary are considered. The most known expert systems for prediction of xenobiotic metabolites are MetabolExpert,11 META,12 Meteor,13 UM-PPS,14 SyGMa, 15 TIMES,16 and JChem Metabolizer module.17 Also, there is software (for example, Metaprint2D-React18) that uses fingerprint-based data mining approach for metabolite prediction. Some software performs the toxicity/mutagenicity assessment of metabolites in addition to metabolites’ estimation. For example, the TIMES program uses a quantitative structure− activity relationahip (QSAR) model for estimating the mutagenicity of genotoxic chemicals.16,19 Such preliminary predictions help to assess the biotransformation net for intermediates with adverse drug effects. Received: October 28, 2016 Published: March 27, 2017 638
DOI: 10.1021/acs.jcim.6b00662 J. Chem. Inf. Model. 2017, 57, 638−642
Application Note
Journal of Chemical Information and Modeling In this manuscript we introduce the method of metabolite prediction, which uses dictionaries of biotransformation and is based on preliminary prediction of possible classes of biotransformation reactions with the subsequent prediction of a single atom in the substrate that is modified during the particular biotransformation reaction. We have considered nine reaction classes (aliphatic and aromatic hydroxylation, N- and O-glucuronidation, N-, S- and C-oxidation, N- and Odealkylation), that are catalyzed by five human cytochromes P450s (1A2, 2C19, 2C9, 2D6, 3A4) and by human UDPglucuronosyltransferase without differentiation into isoforms. The metabolic simulation was integrated with the prediction of acute rat toxicity (LD50 values for intravenous route of administration) and freely available on the web at http://www. way2drug.com/mg.
Figure 1. Schematic representation of xenobiotic metabolism pathway.
The probability of generation of metabolite depends on probabilities calculated for its substrate (from which metabolite is formed) and also on branches leading to it. The probability of parent compound equals to one, P0 = 1, and for each of the layers the sum of probabilities of all metabolites also equals to one. The normalized formula of the probability Pk of metabolite formation is the following:
■
MATERIALS AND METHODS Metabolite Generation. The prediction of substrate modification is based on the rules of transformation. Nine classes of reactions are considered. The process of structural modification is realized through modification of surrounding of single atoms. In the case of hydroxylation, oxidation and glucuronidation reactions it is the process of joining the hydroxyl/carbonyl/carboxyl groups or glucuronic acid component to the considered atom. In the case of dealkylation reaction, it is the process of breaking of atomic bonds. This procedure of the metabolite formation prediction includes three steps. The first step is a prediction of the class of reaction for the entire molecule. This prediction is based on the algorithm of PASS software, which uses substructural atom-centered MNA (multilevel neighborhoods of atoms20) descriptors for representation of the molecular structures and Bayesian-like approach for the building of the SAR classification models.21 The two values are obtained for each of the predicted classes: PR (probability “to be the real reaction”) and PI (probability “to be the impossible reaction”). PR estimates the chance that this predicted class of reaction will be occurred in the experiment. PI estimates the chance that this predicted class of reaction will not be occurred in the experiment. The second step is a prediction of a “reacting atom”, corresponding to a single atom in the substrate that is modified during the particular biotransformation reaction.22 This prediction is based on the SOMP method.23 Two values are calculated for each reaction class for each of the atoms in the molecule: PT (probability “to be true reacting atom”) and PF (probability “to be false reacting atom”). PT estimates the chance that the considered atom will be modified during the considered biotransformation reaction, PF estimates the chance that the considered atom will not be modified during the considered biotransformation reaction. The third step is to calculate the probability of formation of a metabolite. The following probability estimation is calculated for each branch of the xenobiotic metabolism pathway: PC =
Pk =
∑s , c PP s sck ∑s , c , m PP s scm
where Ps is the probability of substrate s (from which metabolite is formed), Psck is the probability of branches c from substrate s to the metabolite k. The summation is taken over all metabolites of the considered layer. Acute Toxicity Calculation. The LD50 values of acute rat toxicity are calculated for a parent compound and each of the generated metabolites. Acute toxicity is considered as the adverse effects occurring within a given time, following a single exposure to a substance.24 LD50 value is one of the important characteristics of acute toxicity that corresponds to the dose causing 50% mortality within 24 h of administration. We use a consensus QSAR model,25 which was created earlier by GUSAR program25−27 using as the training set 920 compounds from Accelrys Toxicity Database (formerly SYMYX MDL Toxicity Database, 2011), for calculation of the log10 representation of LD50 values (mmol/kg) of rats with the intravenous type of administration. The accuracy of this QSAR model was evaluated earlier on the external validation set from 394 compounds. It was shown that R2test and RMSEtest were 0.63 and 0.62, respectively.24 The predicted LD50 values in log10(mmol/kg) are transformed to milligram per kilogram values, which are commonly used by toxicologists. GUSAR uses self-consistent regression (SCR),28 which is based on the regularized least-squares method, QNA (quantitative neighborhoods of atoms)26 and MNA (multilevel neighborhoods of atoms)20 descriptors for the creation of QSAR models. Three methods for estimation of applicability domain (AD) of the QSAR models are used in GUSAR during the prediction of acute toxicity:28 • similarity: the pairwise distance between a query compound and three nearest neighbors in the training set is calculated by Pearson’s correlation coefficient in the space of the independent variables obtained after SCR (in AD if Pearson’s correlation coefficient < 0.7) • leverage (in AD if leverage for a compound < 99th percentile in the distribution of the leverage values calculated for the training set) • assessment of accuracy: the error of prediction for the three most similar compounds in the training set relative
PR PT PR PT + PIPF
It should be noted that different substrates could be transformed to the same metabolite. For example, both substrate 1A and 1B form metabolite 2B (Figure 1). 639
DOI: 10.1021/acs.jcim.6b00662 J. Chem. Inf. Model. 2017, 57, 638−642
Application Note
Journal of Chemical Information and Modeling
biotransformation reactions. Structures of substrates were represented by MNA descriptors.20 Training sets of the second type for prediction of reacting atoms were created by joining of positive and negative SoLA (Structure with one Labeled Atom), presented by LMNA descriptors.8 A detailed description of training set creation is published elsewhere.22 Training procedures and predictions were performed by PASS software with two types of training sets. A description of the training sets and results of leave-one-out cross-validation training procedure is presented in Table 1. The number of compounds for the particular biotransformation reaction is the same for the training sets of both types. The number of positive examples in the training set of the first type equals the number of compounds. In the training sets of the second type, the number of positive examples is the number of positive SoLAs. The invariant accuracy of prediction (IAP) criterion, which is an equivalent of AUC29 (the area under the ROC curve), was used for the estimation of the accuracy during the training procedure. Nc is the number of compounds in the training set (as the same substrates may undergo different reaction classes total number is lower than the sum of Nc values); Np is the number of positive SoLAs; IAP1 is the IAP value calculated for the biotransformation classes prediction; IAP2 is the IAP value calculated for the reacting atoms prediction for each of the biotransformation classes. MetaTox Web Application. The web application for metabolite generation and acute rat toxicity estimation for drug-like compounds was created. It is freely available at the following address: http://way2drug.com/mg. Marvin JS chemical editor 30 is used for input and visualization of molecular structure. Prediction can be made
to the whole training set (in AD if RMSE3NN/RMSEtrain < 1). A freely available web application for prediction of LD50 values for rats with four types of administration was earlier developed and included above-mentioned consensus QSAR model (http://www.way2drug.com/gusar/acutoxpredict.html).
■
RESULTS AND DISCUSSION Data Sets. We have used information about biotransformation presented in the Biovia Metabolite database. The Table 1. Training Set Characteristics biotransformation reaction classes
Nc
Np
IAP1
IAP2
aliphatic hydroxylation aromatic hydroxylation C-oxidation N-oxidation S-oxidation N-glucuronidation O-glucuronidation N-dealkylation O-dealkylation
392 299 69 115 93 320 2264 401 280
508 430 69 121 96 330 2555 422 305
0.81 0.79 0.80 0.87 0.95 0.89 0.89 0.92 0.90
0.91 0.92 0.86 0.99 0.99 0.99 0.99 0.99 0.99
biotransformations that are catalyzed by human enzymes of Phase I and Phase II of dug metabolism and belong to nine reaction classes (aliphatic and aromatic hydroxylation, N- and O-glucuronidation, N-, S- and C-oxidation, and N- and Odealkylation) were selected. The training set of the first type, consisting more than of 3500 substrates of cytochromes P450 and UDP-glucuronosyltransferases was created for preliminary prediction of
Figure 2. Main interface of MetaTox. 640
DOI: 10.1021/acs.jcim.6b00662 J. Chem. Inf. Model. 2017, 57, 638−642
Application Note
Journal of Chemical Information and Modeling Table 2. Toxicity Classes, Used in MetaTox Application toxicity class
1
2
3
4
5
low toxic
interval of LD50 (mg/kg) color
≤0.7 red
(0.7:7] purple
(7:40] orange
(40:300] yellow
(300:700] blue
>700 green
■
ABBREVIATIONS PASS, prediction of activity spectra for substances; MNA, multilevel neighborhoods of atom; LMNA, labeled multilevel neighborhoods of atom; SOM, site of metabolism; SoLA, structure with one labeled atom; LOO CV, leave-one-out crossvalidation; IAP, invariant accuracy of prediction; AUC, area under the roc curve
for single-component uncharged low molecular weight organic molecules with at least three carbon atoms. The user should specify the following parameters (Figure 2): Value (cutoff)all metabolites with Pm higher than specified would be shown. Layer countthe user could specify how many layers (generation) would be generated. Reactionthe user could specify metabolite(s) of which reaction would be shown. The user obtains the results in a few minutes after clicking the button “Send”. In each layer, structures are sorted by probability Pk of metabolite obtaining (see section Metabolite Generation). Examples of generated metabolites with the comparison with known xenobiotic metabolism pathways from ChEMBL31 database (v. 22) can be found in the “Examples” section of the web site. The user can modify the generated xenobiotics metabolism pathway by addition of a new metabolite or by removing the generated metabolite. During the addition of the new metabolite, the user can input molecule by Marvin JS (Figure 2). For each metabolite from a biotransformation pathway, LD50 values (mg/kg) for the rats with the intravenous type of administration is calculated. There is a heat map near each metabolite, which shows a toxicity class according to the toxicity ranking.32 We have used the following colors to distinguish the toxicity classes (Table 2). The user can identify the most toxic metabolite and download SD file of the xenobiotics metabolism pathway by clicking the corresponding buttons.
■
(1) Kirchmair, J.; Howlett, A.; Peironcely, J.; Murrell, D.; Williamson, M.; Adams, S.; Hankemeier, T.; van Buren, L.; Duchateau, G.; Klaffke, W.; Glen, R. How Do Metabolites Differ from Their Parent Molecules and How are they excreted? J. Chem. Inf. Model. 2013, 53, 354−367. (2) Guengerich, F. Cytochrome P450 and Chemical Toxicology. Chem. Res. Toxicol. 2008, 21, 70−83. (3) Guengerich, F. Intersection of Roles of Cytochrome P450 Enzymes with Xenobiotic and Endogenous Substrates. Relevance to Toxicity and Drug Interactions. Chem. Res. Toxicol. 2017, 30, 2−12. (4) Liu, W.; Shi, J.; Zhu, L.; Dong, L.; Luo, F.; Zhao, M.; Wang, Y.; Hu, M.; Lu, L.; Liu, Z. Reductive metabolism of oxymatrine is catalyzed by microsomal CYP3A4. Drug Des., Dev. Ther. 2015, 9, 5771−5783. (5) D’Agostino, J.; Zhang, H.; Kenaan, C.; Hollenberg, P. Mechanism-Based Inactivation of Human Cytochrome P450 2B6 by Chlorpyrifos. Chem. Res. Toxicol. 2015, 28, 1484−95. (6) Rydberg, P.; Gloriam, D.; Zaretzki, J.; Breneman, C.; Olsen, L. SMART-Cyp: A 2D Method for Prediction of Cytochrome P450Mediated Drug Metabolism. ACS Med. Chem. Lett. 2010, 1, 96−100. (7) Zaretzki, J.; Bergeron, C.; Rydberg, P.; Huang, T.; Bennett, K.; Breneman, C. RS-predictor: A new tool for predicting sites of cytochrome P450-mediated metabolism applied to CYP 3A4. J. Chem. Inf. Model. 2011, 51, 1667−1689. (8) Rudik, A.; Dmitriev, A.; Lagunin, A.; Filimonov, D.; Poroikov, V. Metabolism site prediction based on xenobiotic structural formulas and PASS prediction algorithm. J. Chem. Inf. Model. 2014, 54, 498− 507. (9) Dang, N.; Hughes, T.; Krishnamurthy, V.; Swamidass, S. A simple model predicts UGT-mediated metabolism. Bioinformatics 2016, 32, 3183−3189. (10) Kirchmair, J.; Williamson, M.; Tyzack, J.; Tan, L.; Bond, P.; Bender, A.; Glen, R. Computational prediction of metabolism: Sites, products, SAR, P450 enzyme dynamics, and mechanisms. J. Chem. Inf. Model. 2012, 52, 617−648. (11) Darvas, F.. In QSAR in Environmental Toxicology−II; Kaiser, K. L. E., Ed.; Springer: Dordrecht, 1987; p 71. (12) Klopman, G.; Dimayuga, M.; Talafous, J. META. 1. A program for the evaluation of metabolic transformation of chemicals. J. Chem. Inf. Model. 1994, 34, 1320−1325. (13) Marchant, C.; Briggs, K.; Long, A. In silico tools for sharing data and knowledge on toxicity and metabolism: derek for windows, meteor, and vitic. Toxicol. Mech. Methods 2008, 18, 177. (14) Gao, J.; Ellis, L.; Wackett, L. The University of Minnesota Biocatalysis/Biodegradation Database: improving public access. Nucleic Acids Res. 2010, 38, 488−491. (15) Ridder, L.; Wagener, M. SyGMa: combining expert knowledge and empirical scoring in the prediction of metabolites. ChemMedChem 2008, 3, 821. (16) Mekenyan, O.; Dimitrov, S.; Pavlov, T.; Veith, G. A systematic approach to simulating metabolism in computational toxicology. I. The TIMES heuristic modelling framework. Curr. Pharm. Des. 2004, 10, 1273−1293.
■
CONCLUSIONS A new freely available web application for generation and modification of xenobiotics metabolism pathways was created. The user could define different parameters of generation. The probability of a generated metabolite is calculated by using integrated assessment of preliminary prediction of biotransformation classes and prediction of the reacting atom of each of biotransformation classes. MetaTox application is the first web application where metabolism pathway generation is integrated with the prediction of acute toxicity. The web site will be useful in the early drug discovery process.
■
REFERENCES
AUTHOR INFORMATION
Corresponding Author
*E-mail:
[email protected]. ORCID
Anastasia V. Rudik: 0000-0002-8916-9675 Funding
The project was supported by Russian Science Foundation grant 14-15-00449. Notes
The authors declare no competing financial interest. 641
DOI: 10.1021/acs.jcim.6b00662 J. Chem. Inf. Model. 2017, 57, 638−642
Application Note
Journal of Chemical Information and Modeling (17) https://www.chemaxon.com/products/online-tryouts/ metabolizer/ (accessed February 1, 2016). (18) http://www-metaprint2d.ch.cam.ac.uk/metaprint2d-react (accessed February 1, 2017). (19) Mekenyan, O.; Dimitrov, S.; Serafimova, R.; Thompson, E.; Kotov, S.; Dimitrova, N.; Walker, J. Identification of the structural requirements for mutagenicity by incorporating molecular flexibility and metabolic activation of chemicals I: TA100 model. Chem. Res. Toxicol. 2004, 17, 753−66. (20) Filimonov, D.; Poroikov, V.; Borodina, Y.; Gloriozova, T. Chemical similarity assessment through multilevel neighborhoods of atoms: definition and comparison with the other descriptors. J. Chem. Inf. Comput. Sci. 1999, 39 (4), 666−670. (21) Poroikov, V.; Filimonov, D.; Borodina, Yu.; Lagunin, A.; Kos, A. Robustness of biological activity spectra predicting by computer program PASS for non-congeneric sets of chemical compounds. J. Chem. Inf. Comput. Sci. 2000, 40 (6), 1349−1355. (22) Rudik, A.; Dmitriev, A.; Lagunin, A.; Filimonov, D.; Poroikov, V. Prediction of reacting atoms for the major biotransformation reactions of organic xenobiotics. J. Cheminf. 2016, 8, 68. (23) Rudik, A.; Dmitriev, A.; Lagunin, A.; Filimonov, D.; Poroikov, V. SOMP: Web server for in silico prediction of sites of metabolism for drug-like compounds. Bioinformatics 2015, 31, 2046−2048. (24) Nielsen, E.; Ostergaard, G.; Larsen, J. Toxicological Risk Assessment of Chemicals: A Practical Guide; Informa HealthCare: New York, 2008. (25) Lagunin, A.; Zakharov, A.; Filimonov, D.; Poroikov, V. QSAR Modelling of Rat Acute Toxicity on the Basis of PASS Prediction. Mol. Inf. 2011, 30, 241−250. (26) Filimonov, D.; Zakharov, A.; Lagunin, A.; Poroikov, V. QNAbased ‘Star Track’ QSAR approach. SAR QSAR Environ. Res. 2009, 20, 679−709. (27) Lagunin, A.; Zakharov, A.; Filimonov, D.; Poroikov, V. A new approach to QSAR modeling of acute toxicity. SAR QSAR Environ. Res. 2007, 18, 285−298. (28) Zakharov, A.; Varlamova, E.; Lagunin, A. A.; Dmitriev, A. V.; Muratov, E. N.; Fourches, D.; Kuz’min, V. E.; Poroikov, V. V.; Tropsha, A.; Nicklaus, M. C. QSAR Modeling and Prediction of DrugDrug Interactions. Mol. Pharmaceutics 2016, 13 (2), 545−556. (29) Swets, J. Measuring the accuracy of diagnostic systems. Review. Science 1988, 240, 1285−1293. (30) https://www.chemaxon.com/products/marvin/marvin-js/ (accessed February 1, 2017). (31) https://www.ebi.ac.uk/chembl/ (accessed February 1, 2017). (32) Berezovskaya, I. Classification of Substances with Respect to Acute Toxicity for Parenteral Administration. Pharm. Chem. J. 2003, 37, 139.
642
DOI: 10.1021/acs.jcim.6b00662 J. Chem. Inf. Model. 2017, 57, 638−642
