In Silico Quantitative Structure Toxicity Relationship of Chemical Compounds: Some Case Studies

Transcription

1 Send Orders of Reprints at Current Drug Safety, 2012, 7, In Silico Quantitative Structure Toxicity Relationship of Chemical Compounds: Some Case Studies Omar Deeb *, 1 and Mohammad Goodarzi 2 1 Faculty of Pharmacy, Al-Quds University, PO Box 20002, Jerusalem, Palestine 2 Young Researchers Club, Islamic Azad University, Arak Branch, P.O. Box Arak, Markazi, Iran Abstract: Undesirable toxicity is still a major block in the drug discovery process. Obviously, capable techniques that identify poor effects at a very early stage of product development and provide reasonable toxicity estimates for the huge number of untested compounds are needed. In silico techniques are very useful for this purpose, because of their advantage in reducing time and cost. These case studies give the description of in silico validation techniques and applied modeling methods for the prediction of toxicity of chemical compounds. In silico toxicity prediction techniques can be classified into two categories: Molecular Modeling and methods that derive predictions from experimental data. Molecular modeling is a computational approach to mimic the behavior of molecules, from small molecules (e.g. in conformational analysis) to biomolecules. But the same approaches can also be applied for toxicological purposes, if the mechanism is receptor mediated. Quantitative Structure-Toxicity Relationships (QSTRs) models are typical examples for the prediction of toxicity which relates variations in the molecular structures to toxicity. There are many applied modeling techniques in QSTR such as Partial Least Squares, Artificial Neural Networks, and Principal Component Regression (PCR). The applicability of these techniques in predictive toxicology will be discussed with different examples of sets of chemical compounds. Keywords: Drug discovery, in silico methods, molecular modeling, quantitative structure toxicity relationship (QSTR). INTRODUCTION Usually, the exploration of new drugs has concentrated on the required activity, taking into account the bioavailability. The toxicity is left for later stages in the development process. There exist various examples of drugs that have had to be withdrawn because of undesirable toxicity, both in medical trials stages and after reaching the market. Such failures are very expensive because a usual drug takes years, and costs up to $500 million, to reach the market. It is clearly important to discover potential toxicity as soon as possible. For that reason, in recent years pharmaceutical companies have brought toxicity testing, as well as ADME (absorption, distribution, metabolism, excretion) studies, in advance in the drug development process. Predictable toxicity studies try to investigate the effects of the toxicant being tested in laboratory animals exposed to various dosage levels and for different durations. In order to examine the different effects related to various lengths of contact, toxicological studies are generally divided into the following: (a) short-term (acute) toxicity studies (b) chronic toxicity (long term toxicity) studies. A detrimental result of a chemical can be expressed after short term and/or long-term exposure. The short-term (acute) toxicity is most often considered as the concentration that is *Address correspondence to this author at the Faculty of Pharmacy, Al-Quds University, PO Box 20002, Jerusalem, Palestine; Tel/Fax: ; deeb20000il@yahoo.com; deeb.omar@gmail.com lethal to 50% of the organisms (Lethal Concentration, LC50) or causes a measurable effect to 50% of the test organisms (Effective Concentration, EC50). In 2006, the European Council and European Parliament adopted legislation for a new chemical management system called REACH (Registration, Evaluation, Authorization and Restriction of Chemicals). One of the goals of REACH is to ensure a high level of human health and the environment, including the promotion of alternative (non-animal) methods for assessment of hazards of substances. Therefore, the ultimate here would be to use computer-based (in silico) methods to predict toxicity even before a drug candidate was synthesized. Non-testing methods or in silico methods will play an important role for indicating the presence or absence of a toxic property, which in certain cases may be a satisfactory replacement to animal testing. This is particularly true for the quantitative structure toxicity relationship (QSTR). In silico methods are a new development in chemical testing that relies on computer simulation or modeling. The term in silico is a new phrase usually used to represent an experimentation performed by computer, and is linked to the more commonly known biological terms in vivo and in vitro. In silico pharmacology or computational pharmacology is a rapidly growing field. It defines the use of this information in the design of computational models or simulations that can be used to make predictions, suggest hypotheses, and ultimately supply discoveries in therapeutics. - /12 $ Bentham Science Publishers

2 290 Current Drug Safety, 2012, Vol. 7, No. 4 Deeb and Goodarzi There are some outstanding introductions to reviews of computational toxicology [1-10]. Additionally, the website of the institute for health and consumer protection (IHCP): ( _EN.pdf) is an excellent resource. Another website related to the review of software tools for toxicity prediction is: ( R_24489_EN.pdf). The same holds true for the US-based International QSAR Foundation ( which is non-profit research organization dedicated exclusively to creating alternative methods for identifying chemical hazards without additional laboratory testing. One of the validation techniques of in silico models is with an external test set. A clear separation between the training set for model development and a test set for validation is desirable. If it is impossible to create an external test set of enough size, it is possible to perform proper validation with simulated test sets which is considered as another validation technique. The recommended technique for small training sets is leave-one-out (LOO) cross-validation. LOO means that one compound is removed from the training set as test case. A model is created with the remaining training set, and the activity of the test case is predicted and compared with its real activity. The same procedure is repeated for all compounds of the training set. If the training set size is too big, it is possible to remove a certain fraction (e.g. 20%) of the training set compounds at one time, this procedure is called leave many out (LMO) cross-validation. TOXICITY DATABASES Using QSTR methods, it is possible in principle to model every feasible toxic endpoint if the training set of enough size and if structural variety and quality exist. A major blockage in the growth of QSTR models is still the restricted public accessibility of high-quality toxicity data [11]. Two main initiatives have been in progress to standardize and increase the accessibility of toxicity data. The Distributed Structure- Searchable Toxicity (DSSTox) Public Database Network ( [12] is an initiative of Ann Richard (US EPA) to provide a decentralized set of standardized structure searchable toxicity databases. The VITIC database is also a structure-searchable database that is sponsored by a number of pharmaceutical and chemical companies. Furthermore, the database of the United States National Library of Medicine ( nlm.nih.gov/) is also a database on toxicology, hazardous chemicals, environmental health and toxic releases. MOLECULAR MODELING IN TOXICOLOGY Molecular modeling is a computational approach to mimic the behavior of molecules, from small molecules (e.g. in conformational analysis) to biomolecules. Molecular modeling has been used principally in pharmaceutical research for discovery and assessment of new lead compounds. However, the same approach can also be applied for toxicological purposes if the mechanism is receptor mediated and the receptor structure is available. A brief review related to Molecular Modeling techniques is given by Kroemer [13]. The different software used in Molecular Modeling can be found at the Protein Data Bank (PDB) website /home/home.do. Molecular Modeling can be used to clarify mechanisms and to calculate receptor-mediated toxicity. Experimental data is used to supply a record of probable targets associated with chemical toxicity. For a number of these targets, structural data is available or can be retrieved from similar structures by the use of homology modeling [14]. The exact interactions between potential toxicants and the known targets may be modeled by the use of "docking" [15]. QUANTITATIVE STRUCTURE ACTIVITY RELAT- IONSHIP (QSAR) IN TOXICOLOGY (QSTR) QSTRs are based on the hypothesis that the structure of a molecule must enclose the features responsible for its physical, chemical and biological properties, and on the capability to characterize the chemical by one, or more, numerical descriptor(s) or properties. The toxicity of substances is governed by their properties, which in turn are determined by their chemical structure. As a result, there are interrelationships between structure, properties, and toxicity. QSTR is a statistically resulting equation that quantitatively describes a molecular property in terms of descriptors of compounds structure. A simple mathematical relationship is established in any kind of QSTR relations as: Toxicity = f (chemical structure or property) The official birthday of QSAR is the 1962 publication of Hansch et al. [16], which established the use of multiple linear regressions as a significant tool for the relationship of biological activity with descriptors of compounds structure. The linear Hansch equation is given by: log(1/c) = k 1 + k 2 + k 3 where C is the concentration of a compound producing a biological response, is the octanol/water partition coefficient and is the Hammett constant. QSTR can be divided into four crucial steps: (a) conversion of structures into descriptors (parameters); (b) descriptor selection to minimize the risk of chance correlations; (c) deriving the relationship between the molecular descriptors and the toxicological data; (d) and validating the QSTR model and assessing its predictivity. The following scheme summarizes these QSTR steps. Quantitative Structure-Toxicity Relationship (QSTR) Models Activity Data (Y) Prediction Set of Compounds QSTR Y = f(xi) Molecular Descriptors (Xi) Interpretation The QSTR models are developed using a variety of chemometric tools such as multiple linear regression (MLR),

3 In Silico Quantitative Structure Toxicity Relationship of Chemical Compounds Current Drug Safety, 2012, Vol. 7, No partial least squares (PLS), artificial neural network (ANN), support vector machine (SVM) and others. In the case of MLR, the general purpose is to quantify the relationship between several independent or predictor variables and a dependent variable. A multilinear model can be represented as: y = X X x k X k where k is the number of independent variables, 1... k, the regression coefficients and y is the dependent variable. Regression coefficients represent the independent contributions of each calculated molecular descriptor. The statistical quality of the developed QSTR models has been examined by different statistical parameters such as determination coefficient (R 2 ), and standard error of estimate(s). QSTR can model the biological data well. The model will be predictive if it is accepted that all the compounds involved should operate by the same mechanism, because the physico-chemical and structural properties used in the QSTR are believed to be a sign of mechanism of action. Normally, the most important aim in developing a QSTR model is that it can be used for predictive purposes. It is consequently important that the statistics given with the QSTR provide a signal of its predictivity. This is achieved by the use of either internal validation (cross validation) or external validation (use of a test-set), as well as data randomization or Y-scrambling. Predictivity are checked by different validation parameters such as predicted residual sum of squares (PRESS), root mean square error of prediction or calibration (RMSEP or RMSEC), and crossvalidayed R 2 (Q 2 ). In more detail, the PRESS (predicted residual sum of squares) statistic appears to be an important parameter accounting for a good estimate of the real predictive error of the models. Its small value indicates that the model predicts better than chance and can be considered statistically significant. n PRESS = (Y obs - Y pred ) 2 i=1 Root mean square error of prediction or calibration (RMSEP or RMSEC) is a measurement of the average difference between experimental and predicted values at the prediction step. RMSEP can be interpreted as the average prediction error, expressed in the same units as the original response values. The RMSEP was obtained by the following formula: n RMSEP (or RMSEC) = [1/n (Y obs - Y pred ) 2 ] 0.5 i=1 where n is number of compounds. For the leave one out (LOO) cross-validation technique (internal validation), the coefficient of determination (R 2 ) was reported as a measure of the total variance of the response explained by the regression models. In this technique, one compound is eliminated from the data set at random in each cycle, and the model is built using the rest of the compounds. The model thus formed is used for predicting the activity of the eliminated compound. The process is repeated until all the compounds are eliminated once. On the basis of the predicting ability of the model, the cross-validated R 2 (Q 2 ) for the model is determined. The higher is the value of Q 2 (more than 0.5), the better is the model predictivity. The cross-validated R 2 is expressed as: (Y Q 2 obs Y pred ) 2 = 1 (Y obs Y ) 2 where, Y obs and Y pred are the observed and predicted response values respectively and Y is the mean observed value. For the external validation (use of a test set), models are created based on training set compounds and the predictive capacity of the models is judged based on the predictive R 2 (R 2 pred) or Q 2 ext values calculated according to the following equation: 2 R pred (Y obs(test ) Y pred(test ) ) 2 = 1 (Y obs(test ) Y training ) 2 In the above equation, Y pred(test) and Y obs(test) indicate predicted and observed response values with respect to the test set compounds, and Y training indicates the mean response value of the training set compounds. The value of R 2 pred for an acceptable model should be more than 0.5. Another aspect of regression QSTR model validation is the applicability domain. It is a region in chemical space which is defined by the model descriptors and modeled response. It indicates that only those molecules that fall within this domain of applicability can be considered consistent for prediction. In summary, QSTR models should be validated according to OECD (Organization for Economic Cooperation and Development) principles for reliable and acceptable predictions. These principles are: (i) certify clarity in the endpoint being predicted by a given model, (ii) guarantee transparency in the model algorithm that generates predictions of an endpoint, (iii) define an applicability domain, (iv) validate a model by the internal validation (goodness of fit and robustness) as well by external prediction, (v) do not reject models that have no apparent mechanistic basis. More validation tests developed by Roy et al. [42-45] can be used to validate QSTR models. The QSTR for toxicity prediction has been reviewed several times, including in the book edited by Karcher and Devillers [17]. The limiting issue in the development of toxicological QSTRs is the accessibility of high quality experimental toxicity data. Toxicity data used in QSTR evaluations are either safe from the literature or generated particularly for QSTR-type analyses. In the first case, these data, frequently consisting of congeneric series, are limited to neutral organic compounds. In the second case, efforts have often been made to guarantee structural diversity even within a chemical group. A structure- toxicity model is defined and restricted by the character and quality of the data used in model development, and should be applied merely within the model s applicability domain. The perfect QSTR should consist of enough number of compounds for

4 292 Current Drug Safety, 2012, Vol. 7, No. 4 Deeb and Goodarzi satisfactory statistical representation, have a wide range of toxic potency and yield to mechanistic analysis. There are many reviews and papers in the literature that deal with QSAR in toxicology; they use the term QSTR which is quantitative structure toxicity relationship. QSARs for toxicity have been developed for over a century, but the past two decades have seen great development in the area. Toxicological-based QSARs have been developed for a wide variety of end-points including acute and aquatic toxicity, receptor-based toxicities, and human health effects. In this review, we will discuss these issues. Schultz et al., [18] described the history of the use of QSARs in toxicology, for environmental as well as human health effects. An emphasis is made on the science as an answer to the United States Toxic Substance Control Act of In particular, the fundamental concepts and objectives of QSARs for toxicity are reviewed. QSARs for environmental and human health effects are discussed independently. Environmental (ecotoxicity) QSARs have payed attention to modeling congeneric series and nonspecific effects in aquatic organisms through the use of logp (1-octanol/water partition coefficient) to explain hydrophobicity. Molecules that do not fit these QSTRs have been discussed by differences in mechanism of acute toxicity, especially as a result of electro(nucleo)philic relations. Therefore, mechanisms of acute toxicity are also explained. QSAR methods to receptor-mediated effects, such as those exhibited by environmental estrogens and competitive binding to the estrogen receptor, are different from those typically applied to model acute toxic endpoints. Several of these methods, including 3-D QSAR techniques, are reviewed. For human health effects, a local effect such as skin sensitization is often modeled by multivariate QSAR methods such as linear regression and discriminant analysis. The prediction of systemic effects such as carcinogenesis requires consideration of the endpoint and a more mechanistic starting point for modeling. Approaches to predict these endpoints are discussed. One more paper in 2003 by Schultz et al., [19] described the present status of QSAR in toxicology, regarding both environmental and human health effects. Description of environmental QSARs concentrates on current information that relates to the partition of effects anchored in modes of toxic action. Exact consideration is given to the responsesurface method to model the toxic potency of baseline and non-specific soft electrophiles, as well as the progress of rules-based expert systems to assist in the selection of the most appropriate QSAR. Also, a description of the more recent application algorithms such as artificial neural networks to environmental data is discussed. Modern QSAR modeling of receptor-mediated effects such as estrogenicity are described. Furthermore, the current status of modeling human health effects such as carcinogenesis, mutagenesis and others are discussed. We can conclude from this study that environmental endpoints and aquatic toxicity in particular, have been subject to much QSTR analyses. There is no doubt that we can predict the acute environmental toxicity of the majority of chemicals reasonably well. The problem lies in the prediction of the rest of the compounds, which are acting by more specific mechanisms of action. A number of methods are now available to screen compounds for the capability of eliciting a response and for predicting potency. Dearden [20] describes the in silico prediction of drug toxicity. He concentrates on the issue of determining the potential toxicity problems as early as possible in order to minimize drug failures due to toxicity being found in late stages or even in clinical trials. Due to the huge libraries of molecules now being handled by combinatorial chemistry, identification of assumed toxicity is desirable even before synthesis. As a result, the use of predictive toxicology is essential. A number of in silico method for the prediction of toxicity are explained. Quantitative structure-toxicity relationships (QSTRs) have been used for this reason, and are present for a wide range of toxicity endpoints. However, QSTRs are of this classification, they predict whether or not a compound is toxic, and do not give an indication of the degree of toxicity. Examples of all of these are given in this paper. A number of expert systems are available for toxicity prediction, most of them covering a range of toxicity endpoints. We can conclude that there is enormous potential for additional QSTR modeling of toxicity endpoints, as long as adequate experimental data are obtainable for the expansion of the QSTRs. Cronin et al., [21] wrote a review related to the use of in silico methods to predict the chemistry of molecular initiating events related to drug toxicity. The authors explained that drug toxicity pathways can be enormously complex and difficult to totally recognize. Nevertheless, understanding precise parts of the pathway may be simpler. Each toxicity pathway starts with a molecular initiating event (MIE). If an MIE is well understood then it becomes probable to predict which molecules can participate in that particular MIE using in silico methods. In the Cronin et al. review, it is clear that much research has been conducted in trying to understand the exact processes by which drugs cause toxicity, and then in transferring this information to similar molecules. Nevertheless, the mission of understanding toxicity pathways is not an easy one. There are several stages of adverse outcome pathways (AOPs), and understanding how and why each stage happens is complicated. It is clear that the MIE is likely to be related to the properties of a molecule but its effects are dependent on the physiological state of the biological medium. The input information when deciding if a molecule could be toxic is its ability to cause an MIE. The Cronin et al. review has clarified the in silico methods which are presently available for predicting whether molecules are likely to take part in a chemistry-related MIE. Even though these methods can be used to demonstrate probable toxic effects, to understand if a potential MIE will lead to toxicity, a molecule must be compared with other similar molecules with recognized toxicity profiles. This is based on the principle that the adequately comparable molecules will follow the same AOP. Therefore, this comparison is significant because not all molecules with the power to cause an MIE will cause toxicity. Worth et al., [22] paper reviews the quantitative structure activity relationships for acute aquatic toxicity, as well as different methods described in the literature for calculating the aquatic toxicity of chemical substances. It is based on an extended review performed by the European Chemicals

5 In Silico Quantitative Structure Toxicity Relationship of Chemical Compounds Current Drug Safety, 2012, Vol. 7, No Bureau of the European Commission's Joint Research Centre in support of the development of technical guidance for the implementation of the REACH (Registration, Evaluation, Authorization and Restriction of Chemicals) legislation. The paper includes a review of: (a) QSTRs for acute aquatic toxicity by chemical class, (b) QSTRs for acute aquatic toxicity by mode of action, (c) statistically derived QSTRs, (d) structural alerts for excess aquatic toxicity and (e) expert systems that combine structural rules and multiple QSTR models to predict aquatic toxicological endpoints. It is clear from the Worth et al. paper that the QSTR modeling of aquatic toxicity is an exhaustive field of research, in terms of developing predictive potential and advancing the modeling methodology. One restriction in modeling comes from accessibility of high quality data. Existing models were developed from accessible data trying to model a given chemical class or mode/mechanism of action. Cho et al., [23] discussed the in silico quantitative structure toxicity relationship of aromatic nitro compounds. Small compounds often have toxicities that are a function of molecular structural features. Slight variations in structural features can make a large difference in such toxicity. Consequently, in silico methods may be used to show a relationship between molecular toxicities with their structural features. One hundred and forty-eight aromatic nitro compounds were studied. These compounds have similar structures but the activities were determined against different targets with different assays and so consequently they were reported in nine different sets. The authors developed ligand based 2D quantitative structure-toxicity relationship models using twenty chosen topological descriptors. The topological descriptors have several advantages, such as conformational independency and simplistic and less time-consuming computation to yield good results. Multiple linear regression (MLR) analysis was used to correlate variations of toxicity with molecular properties. The following descriptors were identified as fundamental descriptors for different kinds of toxicity: the information index on molecular size, lopping centric index and Kier flexibility index. Also, molecular size, branching and molecular flexibility might be particularly important factors in quantitative structure-toxicity relationship analysis. The Cho et al. study discovered that topological descriptorguided quantitative structure-toxicity relationship provided a very helpful, in silico tool for describing small-compound toxicities. The parent structure of nitrobenzene is shown below: O O - N + As an example, we pick the seventh set nitro compounds with their observed binding affinity to Human lymphocytes. The data indicate that the molecular size index (ISIZ), Kier flexibility index (PHI) and lopping centric index (Lop) with positive coefficients are favorable for high toxicity. The predicted toxicities model is given by the following equation: PT = 0.038ISIZ( ±0.017) PHI (± 0.513) Lop (± 4.6) N = 17; r 2 cv = 0.81; r 2 = 0.85; SEE = 0.51; F = where N is the number of compounds, r 2 cv is the leave-oneout cross-validation coefficient, and SEE is the standard error of estimate and F is fisher value. Bao et al., [24] have built QSTR model of 42 nitrobenzene compounds based on quantum chemical descriptors calculated by PM3 Hamiltonian. Based on the regression coefficient it was found that the toxicity of nitrobenzenes could be affected by the molecular structure, such as the heat of formation and dipole moment. Actually, they have used all 42 compounds for training and have checked the validation of a model built internally. Although the model shows a squared correlation coefficient of crossvalidation of about 0.893, the model is not validated based on the test set and it is not enough to validate the model internally. However, while the authors claim that the model build has good capability for the prediction of nitrobenzene toxicity to the algae (Scenedesmus obliguus), it is not exactly correct due to the lack of sufficient validations. The best model related to toxicity of nitrobenzenes is given in the following equation: pc = TE Mw E LUMO CCR (E LUMO E HOMO ) μ z Q C H f n = 42, Q 2 = 0.893, R = 0.962, p < , SE = where TE is the total energy (ev), Mw: Molecular weight, E LUMO : the energy of the lowest unoccupied molecular orbital (ev), CCR: Core-core repulsion energy (ev), μ z : Z- axix dipole moment (Debye), Q - C : The largest negative atomic charge on a carbon atom (a.c.u) and H f : standard heat of formation (kcal). Carlsen et al. [25] studied the impact on environmental health by residuals of the rocket fuel 1,1-dimethyl hydrazine (heptyl) and its transformation products. The use of QSAR/QSTR modeling for a preliminary screening of a series of secondary pollutants arising from a primary pollution with residual heptyl (1,1-dimethyl hydrazine) and originating from rocket space activities was investigated, and the land-based rocket launching at the Baikonur Cosmodrome is being used as an illustrative example. A series of the studied compounds is further found to be anaerobically biodegradable. None of the studied compounds appear to be bioaccumulating. It was mentioned that further attention should be given to tri- and tetramethyl hydrazine and 1-formyl 2,2-dimethyl hydrazine as well as to the hydrazones of formaldehyde and acetaldehyde as these five compounds may contribute to the overall environmental toxicity of residual rocket fuel and its transformation products. The general structure of 1,1-dimethyl hydrazine is given in the following scheme: NH 2 N H 3 C CH 3

6 294 Current Drug Safety, 2012, Vol. 7, No. 4 Deeb and Goodarzi Thakur and Thakur [26] developed a QSTR model based on cytotoxic concentration for the set of 19 (tetrahydromidazo [4,5,1-jk][1, 4] benzodizepin- 2(1H)-one) (TIBO) derivatives. It was found that the classical physicochemical properties play a dominating role in the modeling of cytotoxicity and molecular property is most responsible for the cytotoxicity of TIBO derivatives. However, it was concluded that the presence of S atom at Z position and the presence of the halogen atoms have a positive impact on the cytotoxicity of TIBO derivatives. The parent structure of TIBO derivatives used in the study of Thakur and Thakur [26] is shown in the following scheme: X HN Z N The best model is given in the following equation: log 1/C = (±0.017)MV (±0.0063)Pc (±0.1757)I X n = 19, Se = , R = , R 2 A = , F = where MV is the molecular volume, Pc: parachor, I x : indicator variable at x position. According to the production and usage in Japan [27], 82 environmental chemicals were randomly selected to estimate the genotoxicity, based upon a Salmonella genotoxicity test using the umu test and systemic toxicity. They used a neural network modeling technique. The model was built based on artificial neural networks and accounted for about 94% of the variation in the genotoxicity results derived by the umutest. The quantum descriptors were calculated by semiempirical method (AM1) and used as input for ANN. The structure of ANN was included with four neurons in the input layer, five neurons in hidden layer, and one neuron in the output layer. The learning involved 400 training cycles. The optimal moment of stopping the training being detected using a complete 20% leave out cross validation experiment with random selection. However, it is a good choice to optimize an ANN model based on random selection because it is not sure when a set is selected as monitoring homogeneity, in addition to whether the training set covers the monitoring set during training. Serra et al. [28] studied a QSTR model of a diverse set of 448 industrially important aromatic solvents. The 50% growth impairment concentration (ICG50) spans the range to 3.36 log units for the ciliated protozoa Tetrahymena was used as toxicity. Several types of molecular descriptors such as topological, geometrical, electronic, and hybrid geometrical-electronic structural features were calculated for each compound. The simulated annealing technique and a genetic algorithm were used as feature selection techniques to select the most relevant sets of descriptors. The Multiple Linear Regression as linear and Computational Neural Networks (CNNs) as non linear modeling techniques were derived and then used to predict the ICG50 values for an N X' R external set of representative compounds. Based on their study, a common theme between the descriptors chosen from the linear and nonlinear methods was that the toxicity of these compounds to Tetrahymena may potentially be related to the size, degree of branching and hydrogen-bonding ability of the molecule of interest. Senior et al., [29] developed a linear QSTR model based on quantum chemical and topological descriptors of some organophosphorus compounds (OP) and their toxicity LD50 as a dermal. Two descriptors gave higher correlation coefficients and a greater affect on the prediction, namely the Heteroatom Corrected Extended Connectivity Randic index (1XHCEC) and the Density Randic index (1XDen). They have applied the model built on a test set and found that the sulfur atoms in the test set must be replaced with oxygen atoms to achieve improved toxicity. It was concluded that to get high toxic OP compound, the compound must be with a higher charge density distribution. To achieve this, the compound must contain higher electronegative atoms like oxygen atoms instead of sulfur atoms. Roy et al., [30] have been studying QSTR of different organic compounds. They have introduced the extended topochemical atom (ETA) indices which were used as descriptors in QSTR studies. This type of descriptor can be easily calculated from two-dimensional molecular representation and showed that it has the potential to be used in QSTR studies. It was used to developed predictive toxicity models on a large dataset of 459 diverse chemicals against fathead minnow (Pimephales promelas) [30]. A QSTR model developed for herg K + channel blocks activity of diverse functional drugs using different chemometric tools, such as factor analysis followed by multiple linear regression (FA-MLR), stepwise regression and partial least squares [31]. 66 inhibitors were split into training and test sets using the K-means clustering technique. The model built based on ETA descriptors predicted that herg channel blocking increases with the increase of molecular bulk, while electron richness decreases with the increase of functionalities of the carboxylic acid group and the aliphatic tertiary nitrogen fragment. A QSTR model of 77 aromatic aldehydes compounds was built based on Quantum chemical descriptors [32]. Molecular descriptors were calculated in different basis level and 58 models were developed using partial least squares (PLS) and genetic partial least squares (G/PLS). The predictive capacity of the models was examined based on leave-one-out and randomization test predictions for the training set compounds, and model derived predictions for the test set compounds. It was found that the HF/6-31G(d) level show better predictivity (based on overall prediction) than the models developed on any of the other considered levels. Thirty different QSTR models of 77 aromatic aldehydes compounds were built using different chemometric tools. All the models have been validated using internal and external validation techniques [33]. The ETA models have been compared with those developed using various non-eta topological indices. The statistical quality of the ETA models was found to be comparable to that of the non-eta models.

7 In Silico Quantitative Structure Toxicity Relationship of Chemical Compounds Current Drug Safety, 2012, Vol. 7, No Several linear QSTR models were built using different statistical tools such as stepwise regression analysis, multiple linear regression with factor analysis (FA-MLR) as the preprocessing step, partial least squares (PLS) regression and principal component regression (PCR) [34]. The QSTR models were studied on toxicity of 91 organic compounds to Chlorella vulgaris and have been performed using extended topochemical atom (ETA) indices. The results suggest that the ETA descriptors are sufficiently rich in chemical information and have the ability to encode the structural features contributing significantly to the toxicity of organic chemicals to C. vulgaris. A QSTR model was established for 297 structurally diverse chemicals for their toxicity to Daphnia magna using mechanistically interpretable descriptors [35]. 2D and 3D types of descriptors were used as predictor variables. The QSAR models were developed by stepwise multiple linear regression (MLR), partial least squares (PLS), genetic function approximation (GFA), and genetic PLS (G/PLS). All the models were validated internally and externally. The PLS model using seven descriptors was the best model due to internal and external validation results. The PLS model suggests that higher lipophilicity and electrophilicity, less negative charge surface area and presence of ether linkage, hydrogen bond donor groups and acetylenic carbons are responsible for greater toxicity of chemicals. QSTR model was developed for the toxicity of 384 diverse aromatic compounds to Tetrahymena pyriformis using ETA indices [36]. The results obtained by the ETA descriptors were compared with those derived from various non-eta topological descriptors and also the combined set of descriptors encompassing the ETA and non-eta parameters. The K-mean clustering technique was used to split the data set into training and test sets. Different statistical analyses (factor analysis followed by multiple linear regression (FA-MLR), stepwise regression and partial least squares (PLS)) were performed with the training set compounds to develop QSTR models using the topological descriptors. Based on the obtained result it was indicated that the use of ETA descriptors with non-eta descriptors improved the statistical quality of the non-eta models. It was found that halogen atoms (hydrophobicity), volume parameter (bulk) and nitrogen containing functionalities (polarity) are important for developing QSTR models for the diverse aromatic compounds. A QSTR model of the acute lethal toxicity of 51 benzene derivatives was developed using ETA descriptors [37]. The different models when using ETA or non-eta (different topological indices (Wiener W, Balaban J, flexibility index, Hosoya Z, Zagreb), molecular connectivity indices, E-state indices and kappa shape indices) and physicochemical parameters (AlogP 98, MolRef, H_bond_donor and H_bond_acceptor ) descriptors were compared. Genetic function approximation (GFA) and factor analysis (FA) were used as the data-preprocessing steps for the development of final multiple linear regression (MLR) equations. Principalcomponent regression (PCR) was also used to extract the total information from the ETA/non-ETA/combined matrices. All the models developed were cross-validated using leave-oneout (LOO) and leave-many-out techniques. The model based on ETA showed that the toxicity of studied compounds depended on molecular size. Furthermore, the toxicity increases with functionality contribution of chloro substituent and decreases with those of methoxy, hydroxy, carboxy and amino groups. The toxicity of a set of 56 phenylsulfonyl carboxylates to Vibrio fischeri was modeled based on ETA indices and using genetic function approximation (GFA) as the statistical tool. However, again a comparison between those models using of ETA and non-eta was made [38]. In the above mentioned studies, the authors compare the different models produced by using ETA descriptors, non- ETA descriptors and finally combined ETA and non-eta descriptors. This is done by using different chemometric tools. For example, the best model obtained in [36] by using ETA descriptors in the case of PLS analysis for the toxicity data is given by the following equation: Log(IGC 50 ) -1 = [ ' F] Cl [ ' F] Br [ ' F] F [ ' F] NO [ ' F] NH2 N Training = 288; Q 2 = 0.740; R 2 = 0.752; R = 0.867; R 2 a= 0.749; F = (df3, 284); PRESS = ; N Test = 96; R 2 pred = where is Sum of values of all non-hydrogen vertices of a molecule, [ ' F] Cl is the functionality contribution for the chlorine atom, 2.038[ ' F] Br is the functionality contribution for the bromine atom, [ ' F] F is the functionality contribution for the fluorine atom, [ ' F] NO2 is the functionality contribution for the nitro group and [ ' F] NH2 is the functionality contribution for the amino group. Roy et al., have been applying ETA descriptors on different sets of toxic compounds. Based on the mentioned studies, it can be concluded that ETA descriptors have sufficient power to encode the chemical information that significantly contributes to the toxicity of different types of compounds under study. Recently, Roy et al., [39] published a paper related to development of classification and regression based QSAR models to predict rodent carcinogenic potency using the oral slope factor. In this paper, the authors picked carcinogenicity as toxicological endpoints posing the highest fear for human health. Oral slope factors (OSFs) are used to estimate quantitatively the carcinogenic potency or the risk associated with exposure to the chemical by oral route. Roy et al., have developed quantitative structure-carcinogenicity correlation models for rodents based on the carcinogenic potential of 70 chemicals with wide variety of molecular structures, spanning a large number of chemical classes and biological mechanisms. All the developed models have been assessed according to the Organization for Economic Cooperation and Development (OECD) principles for the validation of QSAR models. They have also attempted to develop a carcinogenicity classification model based on Linear Discriminant Analysis (LDA). The in silico studies make it possible to obtain a QSAR interpretation of the structural information of carcinogenicity along with identification of the discriminant functions between lower and higher carcinogenic compounds by LDA. Pharmacological distribution diagrams (PDDs) are used as a visualizing technique for the identification and selection of chemicals with lower carcinogenicity. From this study, the authors conclude that combined analysis of the descriptors used in the best QSAR equation and

8 296 Current Drug Safety, 2012, Vol. 7, No. 4 Deeb and Goodarzi LDA suggest that the carcinogenicity often depends on the electrophilicity and particular structural fragments of the chemicals. These electrophilic compounds can exert their carcinogenic effects through covalent interaction with cellular macromolecules. Based upon the descriptor MAXDP which corresponds to maximal electrotopological positive variation in the best QSAR model, the majority of the training set chemicals could be considered as electrophiles that interact with DNA, RNA and protein macromolecules which majorly act as nucleophiles. One of the major identified fragments for lower carcinogenic property is the nrnnox (a descriptor corresponds to number of N-nitroso groups -aliphatic-) fragment which has a negative contribution to carcinogenicity. Most interestingly, the nrnnox fragment is also identified as the major discriminatory feature between higher and lower carcinogenic group by LDA analysis, and it has a positive involvement towards lower carcinogenic group. It should be noteworthy to mention that all the nitroso compounds have carcinogenic properties, but the presence of the nrnnox index in a particular compound can discriminate between higher and lower carcinogens. Hemmateenejad et al., [40] performed a study in which the effect of the electronic and physicochemical parameters on the carcinogenesis activity of some sulfa drugs using QSAR analysis based on genetic-mlr and genetic-pls was discussed. In this paper, a QSAR study has been applied to a data set of 18 sulfa drugs with carcinogenesis activity. Two types of molecular descriptors, including chemical and electronic, were used to derive a quantitative relation between the carcinogenesis activity and structural properties. Two multi-parametric equations with good statistical qualities were obtained using genetic algorithms multiple linear regression (GA-MLR) methods. In addition, genetic algorithm-partial least squares (GA-PLS) regression was used to model the structure-activity relationships more precisely. The results confirmed the superiority of the results obtained by GA-PLS relative to GA- MLR. The significant effect of the highest occupied molecular orbital energy (HOMO) on the carcinogenic activity was explained in the context of the shape of this orbital. The molecular structure of the sulfa drugs is shown below: H 2 N The best model using GA-PLS is given by the following equation: DFcar = (±0.03) Pol (±3.2) SSPC + 7.9(±1.2) SSC + 3.9(±0.3) E HOMO - 4.0(±0.3) + 3.3(±0.4) MNC - 6.4(±0.7) S - 5.8(±0.8) N = 18, R 2 = , MSE = 0.443, Q 2 LOO = , Q 2 L4O = , F = 48.6 where Pol is polarizability, SSPC is sum of squares of positive charges, SSC is sum of squares of charges, EHOMO is the highest occupied molecular orbital energy, is electrophilicity, MNC is most negative charge, S is softness and is hardness. At the end, we can see that in silico methods and specifically QSTR can be used to detect the toxicity of drug O S O H N R like compounds. An excellent review by Ekins et al., [41] related to in silico pharmacology for drug discovery, discussed the different methods for virtual ligand screening and profiling. Ekins et al., have briefly described the development of in silico pharmacology through the development of methods, including databases, quantitative structure-activity relationships, similarity searching, pharmacophores, homology models and other molecular modelling, machine learning, data mining, network analysis tools and data analysis tools that use a computer. We have introduced how some of these methods can be used for virtual ligand screening and virtual affinity profiling. CONCLUSIONS Current in silico tools are developed in a way to play an important task in the preclinical evaluation of toxic effects. The precision of in silico predictions can be several cases, at least analogous to in vitro and in vivo alternatives, and it is likely that more enhancements are attainable with the combination of biological information. Nevertheless, it is crucial to know the restrictions of in silico methods and to stay away from blindly applying such methods to every case. It is beneficial to use methods that present the justification for their predictions in a definite approach, and that designate the restriction of their predictions obviously (such as molecules that fall beyond the applicability domain of the model). In the case of defective predictions, it is better to carry out additional biological experiments to elucidate these issues than to blindly put faith in silico predictions. The in silico / QSTR fields represent the most promising complementary methods to achieve intelligent testing strategies. The relative contribution of in silico / QSTR to the toxicology will depend on the availability of high quality in vivo data. In summary, in silico tools have a brilliant future in toxicology. They provide the tools to evaluate the toxicity of drug like compounds, and they help to combine various approaches in more intelligent ways than a series of tests. CONFLICT OF INTEREST The authors do not have competing financial interests to declare. ACKNOWLEDGEMENTS Deeb wishes to thank Dr. Andrew Buchanan from Alquds- Bard for his help in editing suggestions of the manuscript. PATIENT S CONSENT Declared none. REFERENCES [1] Durham SK, Pearl GM. Computational methods to predict drug safety liabilities. Curr Opin Drug Discov Devel 2001; 4: [2] Greene N, Naven R. Early toxicity screening strategies. Curr Opin Drug Discov Devel 2009; 12: [3] Greene N. Computer systems for the prediction of toxicity: an update. Adv Drug Deliv Rev 2002; 54:

9 In Silico Quantitative Structure Toxicity Relationship of Chemical Compounds Current Drug Safety, 2012, Vol. 7, No [4] Helma C. In silico predictive toxicology: the state-of the- art and strategies to predict human health effects. Curr Opin Drug Discov Devel 2005; 8, [5] Kavlock RJ, Ankley G, Blancato J, et al. Computational toxicology - a state of the science mini review. Toxicol Sci 2008; 103: [6] Merlot, C. In silico methods for early toxicity assessment. Curr. Opin. Drug Discov Devel 2008; 11: [7] Nigsch F, Macaluso NJ, Mitchell JB, Zmuidinavicius D. Computational toxicology: an overview of the sources of data and of modelling methods. Expert Opin Drug Metab Toxicol 2009; 5: [8] Simon-Hettich B, Rothfuss A, Steger-Hartmann T. Use of computer-assisted prediction of toxic effects of chemical substances. Toxicol 2006; 22: [9] Van de Waterbeemd H. High-throughput and in silico techniques in drug metabolism and pharmacokinetics. Curr Opin Drug Disc Devel 2002; 5: [10] Veith GD. On the nature, evolution and future of quantitative structure-activity relationships (QSAR) in toxicology. SAR and QSAR Env Res 2004; 15: [11] Cronin M. Toxicological information for use in predictive modeling: Quality, sources, and databases. In: Helma C, Ed. Predictive Toxicology, Marcel Dekker, New York (2005). [12] Richard A: DSSTox web site launch: Improving public access to databases for building structure-toxicity prediction models. Preclinica 2004; 2: [13] Kroemer RT. Molecular modelling probes: Docking and scoring. Biochem Soc Trans 2003; 31: [14] Hillisch A, Pineda LF, Hilgenfeld R. Utility of homology models in the drug discovery process. Drug Discov Today 2004; 9: [15] Kuntz ID. Structure-based strategies for drug design and discovery. Science 1992; 257: [16] Hansch C, Maloney PP, Fujita T, Muir RM, Correlation of biological activity of phenoxyacetic acids with Hammett substituent constants and partition coefficients. Nature 1962; 194: [17] Karcher W, Devillers J. SAR and QSAR in environmental chemistry and toxicology: scientific tool or wishful thinking? In: Karcher W, Devillers J. Eds. Practical Applications of Quantitative Structure-Activity Relationships (QSAR) in Environmental Chemistry and Toxicology. Kluwer, Dordrecht. 1990; pp [18] Schultz T, Cronin M, Walker J. Aptula A. Quantitative structureactivity relationships (QSARs) in toxicology: a historical perspective. J Mol Struct (Theochem) 2003; 622: [19] Schultz T, Cronin M, Netzeva T. The present status of QSAR in toxicology. J Mol Struct. (Theochem) 2003; 622: [20] Dearden J. In silico prediction of drug toxicity. J Comp-Aided Mol Design 2003; 17: [21] Ellison C, Enoch S. And Cronin M, A review of the use of in silico methods to predict the chemistry of molecular initiating events related to drug toxicity. Exp Opin Drug Metab Toxicol 2011; 7: [22] Netzeva T, Pavan M. Worth A. Review of (Quantitative) structureactivity relationships for acute aquatic toxicity. QSAR. Comb Sci 2008; 27: [23] Pasha F, Neaz M, Cho S, Ansari M, Mishra S, Tiwari S. In silico quantitative structure-toxicity relationship study of aromatic nito compounds. Chem Biol Drug Des 2009; 73: [24] Bao Y, Huang Q, Li Y, Li N, He T, Feng C. Prediction of nitrobenzene toxicity to the algae (Scenedesmus obliguus) by quantitative structure-toxicity relationship (QSTR) models with quantum chemical descriptors. Environ Toxicol Pharm 2012; 33: [25] Carlsen L, Kenessov B, Batyrbekova S. A QSAR/QSTR study on the environmental health impact by the rocket fuel 1,1-dimethyl hydrazine and its transformation products. Environ Health Insights 2008; 1: [26] Thakur A, Thakur M. Chem-Bioinformatics: QSTR studies on TIBO derivatives. J Biochem Mol Toxicol 2009; 23: [27] Shoji R, Kawakami M. Prediction of genotoxicity of various environmental pollutants by artificial neural network simulation, Mol Divers. 2006; 10: [28] Serra J, Jurs P, Kaiser K. Linear regression and computational neural network prediction of tetrahymena acute toxicity for aromatic compounds from molecular structure. Chem Res Toxicol 2001; 14: [29] Senior S, Madbouly M, El massry AM. QSTR of the toxicity of some organophosphorus compounds by using the quantum chemical and topological descriptors. Chemosphere 2011; 85: [30] Roy K, Narayan Das R. QSTR with extended topochemical atom (ETA) indices. 15. Development of predictive models for toxicity of organic chemicals against fathead minnow using secondgeneration ETA indices. SAR QSAR Environ Res 2012; 23: [31] Roy K, Ghosh G. QSTR with extended topochemical atom (ETA) Indices. 13. Modeling of herg K+ channel blocking activity of diverse functional drugs using different chemometric tools. Mol Simulat 2009; 15: [32] Kar S, Harding AP, Roy K, Popelier P. QSAR with quantum topological molecular similarity indices: toxicity of aromatic aldehydes to Tetrahymena pyriformis. SAR QSAR Environ Res 2010; 21: [33] Roy K, Narayan Das R. QSTR with extended topochemical atom (ETA) indices. 14. QSAR modeling of toxicity of aromatic aldehydes to Tetrahymena pyriformis. J Hazard Mater 2010; 183 : [34] Roy K., Ghosh G. QSTR with extended topochemical atom (ETA) indices. 9. Comparative QSAR for the toxicity of diverse functional organic compounds to Chlorella vulgaris using chemometric tools. Chemosphere 2007; 70: [35] Kar S, Roy K, QSAR modeling of toxicity of diverse organic chemicals to Daphnia magna using 2D and 3D descriptors. J Hazard Mater 2010; 177: [36] Roy K, Ghosh G, QSTR with extended topochemical atom (ETA) indices. 12. QSAR for the toxicity of diverse aromatic compounds to Tetrahymena pyriformis using chemometric tools. Chemosphere 2009; 77: [37] Roy K, Ghosh G. QSTR with extended topochemical atom (ETA) indices. VI. Acute toxicity of benzene derivatives to tadpoles (Rana japonica). J Mol Model 2006; 12: [38] Roy K, Ghosh G. QSTR with extended topochemical atom indices. Part 5. Modeling of the acute toxicity of phenylsulfonyl carboxylates to Vibrio fischeri using genetic function approximation. Bioorg Med Chem 2005; 13: [39] Kar S, Deeb O, Roy K. Development of classification and regression based QSAR models to predict rodent carcinogenic potency using oral slope factor. Ecotoxicol Environ Saf 2012; 82: [40] Deeb O, Hemmateenejad B, Jaber A, Garduno-Juarez R, Miri R. Effect of the electronic and physicochemical parameters on the carcinogenesis activity of some sulfa drugs using QSAR analysis based on genetic-mlr and genetic PLS. Chemosphere 2007; 67: [41] Ekins S, Mestres J, Testa B. In silico pharmacology for drug discovery: methods for virtual ligand screening and profiling. Brit J Pharmacol 2007; 152: [42] Roy PP. Paul S, Mitra I, Roy K. On two novel parameters for validation of predictive QSAR models. Molecules 2009; 14: [43] Roy K, Mitra I, Kar S, Ojha PK. Das RN, Kabir H. Comparative studies on some metrics for external validation of QSPR models. J Chem Inf Model 2012; 52: [44] Ojha PK, Mitra I, Das RN, Roy K. Further exploring r 2 m metrics for validation of QSPR models. Chemom Intell Lab Sys 2011; 107: [45] Mitra I, Saha A, Roy K, Exploring quantitative structure-activity relationship studies of antioxidant phenolic compounds obtained from traditional Chinese medicinal plants. Mol Simul 2010; 36: Received: June 30, 2012 Revised: September 8, 2012 Accepted: September 26, 2012