Multispecies QSAR Modeling for Predicting the Aquatic Toxicity of

---

Article pubs.acs.org/crt

Multispecies QSAR Modeling for Predicting the Aquatic Toxicity of Diverse Organic Chemicals for Regulatory Toxicology Kunwar P. Singh,*,†,‡ Shikha Gupta,†,‡ Anuj Kumar,§ and Dinesh Mohan∥ †

Academy of Scientific and Innovative Research, Anusandhan Bhawan, Rafi Marg, New Delhi-110 001, India Environmental Chemistry Division, CSIR-Indian Institute of Toxicology Research, Post Box 80, Mahatma Gandhi Marg, Lucknow-226 001, India § JCDM College of Pharmacy, Post Box No-81, Barnala Road, Sirsa-125055, India ∥ School of Environmental Sciences, Jawaharlal Nehru University, New Delhi 110067, India ‡

S Supporting Information *

ABSTRACT: The research aims to develop multispecies quantitative structure−activity relationships (QSARs) modeling tools capable of predicting the acute toxicity of diverse chemicals in various Organization for Economic Co-operation and Development (OECD) recommended test species of different trophic levels for regulatory toxicology. Accordingly, the ensemble learning (EL) approach based classification and regression QSAR models, such as decision treeboost (DTB) and decision tree forest (DTF) implementing stochastic gradient boosting and bagging algorithms were developed using the algae (P. subcapitata) experimental toxicity data for chemicals. The EL-QSAR models were successfully applied to predict toxicities of wide groups of chemicals in other test species including algae (S. obliguue), daphnia, fish, and bacteria. Structural diversity of the selected chemicals and those of the end-point toxicity data of five different test species were tested using the Tanimoto similarity index and Kruskal−Wallis (K−W) statistics. Predictive and generalization abilities of the constructed QSAR models were compared using statistical parameters. The developed QSAR models (DTB and DTF) yielded a considerably high classification accuracy in complete data of model building (algae) species (97.82%, 99.01%) and ranged between 92.50%−94.26% and 92.14%−94.12% in four test species, respectively, whereas regression QSAR models (DTB and DTF) rendered high correlation (R2) between the measured and model predicted toxicity end-point values and low mean-squared error in model building (algae) species (0.918, 0.15; 0.905, 0.21) and ranged between 0.575 and 0.672, 0.18−0.51 and 0.605−0.689 and 0.20−0.45 in four different test species. The developed QSAR models exhibited good predictive and generalization abilities in different test species of varied trophic levels and can be used for predicting the toxicities of new chemicals for screening and prioritization of chemicals for regulation.

1. INTRODUCTION

Algae produce oxygen and organic substances for other aquatic organisms. Adverse effects of chemicals on algae may cause oxygen depletion and declined primary productivity of an ecosystem.9 The algal growth inhibition test is used for toxicity assessment of chemicals.10 Reliable algal toxicity data of chemicals may help to reduce the number of fish needed for regulatory toxicity testing.11 In algal toxicity tests, the effective concentration (EC50) of the test chemical is determined through measuring the growth after exposure for 72 h.12 Daphnia is widely used as a standard test organism in aquatic toxicology and allows for the most reproducible results. The 48 h Daphnia acute immobilization test is used for short-term toxicity (EC50) assessment of chemicals. The 96 h LC50 test were used to measure the susceptibility and survival potential of fish to toxic

A large number of industrial chemicals are manufactured and used for various purposes, and many new molecules are added to the existing inventory every day. The regulatory agencies have emphasized the safety assessment of the existing and new molecules prior to their manufacture and use.1 According to REACH (Registration, Evaluation, Authorization, and Restriction of Chemicals), the basic ecotoxicological information requirements for substances manufactured per year include short-term toxicity testing on crustaceans (preferred species Daphnia2), growth inhibition on aquatic plants (preferred species algae3), and toxicity testing on fish.4 The crustaceans, algae, and fish belong to different trophic levels.5 The choice of these three trophic levels (primary producers and primary and secondary consumers) is considered to be relevant in order to protect aquatic ecosystems.6 Furthermore, algae, daphnids, and fish are often used for cross-species comparisons of sensitivity.7,8 © 2014 American Chemical Society

Received: October 6, 2013 Published: April 17, 2014 741

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

Figure 1. Radar chart of the selected descriptors in (a) algae (P. subcapitata), (b) algae (S. obliguue), (c) daphnia, (d) fish, and (e) bacteria data sets.

substances, which is the last chain in the aquatic food cycle.13 The bacterial toxicity assay is another alternative to higher organisms in toxicity testing, and Vibrio f ischeri, a luminescent gramnegative bacterium, is most widely used for the ecotoxicity analysis.14 The bacterial bioassays have advantages as these organisms due to the short life cycle respond promptly to change in environmental conditions. Several toxicological studies revealed a significant correlation between the Vibrio f ischeri test with fish, crustaceans, and algae toxicity bioassays.15 However, the experimental approach for the in vivo toxicity testing of the chemicals is expensive, labor intensive, and timeconsuming.16 Therefore, more reliable computational modeling methods are required for predicting the toxicities of the chemicals. The knowledge of physicochemical properties, fate, and ecotoxicological effects are fundamental for the environmental risk assessment of the chemicals. Several quantitative structure−

activity relationships (QSAR) using a univariate linear regression (LR) approach have been used to study the ecotoxicity (EC50) of some selected categories of organic compounds, such as alcohols,17−19 amines,18−20 phenols,19,21−23 aromatics,17−19,24 chlorophenols,25 anilines,5 triazoles, and (benzo-)triazoles,26,27 narcotics,28−31 nitriles,32 and acetaldehydes.33 While individual class-based QSAR models have been proposed for chemicals, a multiclass model is more difficult to achieve.34 Several QSAR models, such as ECOSAR,35 ASTER, CNN, TOPKAT, OASIS,36 and MCASE37 have been used for toxicity prediction of chemicals; however, they yielded poor correlations37 and still require improvement.38 The local models performed well for a limited chemical domain; these were not applicable to assess a large diverse set of chemical structures,39 and moreover, these studies, in general, consider single test species. Although, some researchers27,40−42 report QSARs for predicting toxicity based on interspecies correlations, no attempts have yet been made to 742

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

2.2. Molecular Descriptors, Feature Selection, and Data Processing. In this study, 174 molecular descriptors (topological, electronic, geometrical, and constitutional) were calculated for each compound using Chemistry Development Kit50 (CDK v 1.0.3). The topological, electronic, geometrical, and constitutional descriptors were calculated by 2D structures of the molecules, taken in the form of SMILES (simplified molecular input line entry system). SMILES for the chemical compounds in P. subcapitata, daphnia, and fish toxicity data were obtained using Chemspider,51 and those in V. f ischeri and S. obliguue data were taken from Zhang et al.49 Since, all of the calculated molecular descriptors may not be appropriate for QSARs analysis, less significant descriptors were removed. For the selection of initial features, model-fitting approaches were considered. The descriptors which exhibited a constant or near constant value and low variation were excluded. EL-based QSAR models were trained by using the set of remaining descriptors, and scoring functions were computed to rank the contribution of features in the current data. The lowest ranked features were then removed.52 Using the remaining set of descriptors, the EL-QSARs were retained, and the prediction accuracies (misclassification rate and mean squared error of prediction) were computed by 10-fold cross-validation. Finally, sets of nine descriptors for classification and eight descriptors for regression were considered here. The distribution of selected descriptors for different data sets is shown in Figure 1. Since the aim of the present study is to develop accurate and reliable multispecies QSAR models for predicting the ecotoxicity of untested chemicals, the constructed QSAR models should be validated/tested using new chemical moieties. For developing the QSAR models, the data were split into training and test sets using a random distribution approach, which is a reliable approach to evaluate model validity.53 Test sets, thus obtained (when defined prior to analysis), are considered as the gold standard to assess real predictivity.54 Categorization of compounds as toxic (EC50 ≤ 100 mg/L) and nontoxic (EC50 > 100 mg/L) was performed according to the EEC criteria.16 The EC50 values were taken as −log EC50 (mmol/L) for regression QSAR modeling. 2.3. Diversity in Data. For global model development,55 the structural diversity of the chemicals was measured by using the Tanimoto similarity index (TSI). In this method, the structure of the chemical compounds to be compared are decomposed into fragments. The TSI ranges from 0 (no similarity) to 1 (pairwise similarity). Smaller TSI means compounds have good diversity.56 A good cutoff for biologically similar molecules52 is 0.7 or 0.8. TSI values for the chemicals in algae (P. subcapitata), algae (S. obliguue), daphnia, fish, and marine bacteria data are 0.052, 0.101, 0.046, 0.057, and 0.089, respectively. These values suggest that the chemical compounds considered in this work represent sufficiently high structural diversity. It warrants model stability, and the external test sets are suitable for assessing the predictive performance of the developed model. Moreover, for robust multispecies QSARs, it is also important to ensure that the experimental end-points in the different toxicity test species considered belong to different distributions. The Kruskal− Wallis (K−W) test was used to analyze the experimental toxicity data pertaining to five different test species considered here. It compares the residual value of each toxicity end-point among different test species with the significance level (p < 0.05) suitable for comparing data sets with uneven sample numbers57 and does not rely on the data normality assumption. The K−W test statistic (H) was computed as follows:58

develop multispecies QSARs (model built using ecotoxicity data in one species and capable of predicting toxicity in other test species). Therefore, global toxicity prediction QSAR models are required to predict the toxicity of diverse chemicals in multiple test species. In recent years, ensemble learning (EL) methods43 have emerged as unbiased tools for modeling the complex relationships between a set of independent and dependent variables and have been applied successfully in various research areas.44 In general, these methods are designed to overcome problems with weak predictors45 and avoid the overfitting of data in training.46 Decision tree forest (DTF) and decision treeboost (DTB) implementing bagging and boosting techniques, respectively, are relatively new methods for predictive modeling.44 These techniques are inherently nonparametric statistical methods and make no assumption on the distribution of the values of the independent variables and can handle highly skewed data.47 To our knowledge, ensemble learning methods have not yet been applied to toxicity prediction modeling. Selection of relevant molecular descriptors in QSAR modeling is an important issue. A variety of descriptors derived using semiempirical and empirical methods based on quantum mechanical calculations have been used earlier.20,23,32 Hence, there is a need to develop toxicologically relevant QSARs using simple properties derived from the chemical’s structure. The present study is focused on the development of multispecies QSAR models for ecotoxicity prediction of diverse chemicals. Here, EL-based classification and regression QSAR models (DTB and DTF) were developed for predicting the toxicity classes and the toxicity end-point of the chemicals. The predictive and generalization abilities of the constructed QSAR models were evaluated using statistical parameters, and the performance of these models was tested with the toxicity data of other test species. Accordingly, we developed EL-QSAR models using algae (Pseudokirchneriella subcapitata) toxicity (EC50) data and successfully applied for toxicity prediction of diverse chemicals in other test species, such as algae (Scenedesmus obliguue), daphnia (Daphnia magna), fish (Oryza latipes), and bacteria (Vibrio f ischeri).

2. MATERIALS AND METHODS 2.1. Data Sets. In this study, toxicity data from multiple sources were considered. The data in the ecotoxicity database were gathered by the Japanese Ministry of Environment and are available online.48 The tests were conducted according to the OECD test guidelines performed under good laboratory practices (GLP). We considered the acute toxicity expressed as the EC50 or LC50 for three aquatic species from three different trophic levels, one alga (P. subcapitata, 72 h EC50 calculated from average specific growth rate), one daphnid (D. magna, 48 h EC50), and one fish (O. latipes, 96 h LC50). The sensitivity of O. latipes is comparable to that of other fish species.6 For developing predictive QSAR models, algae (P. subcapitata) ecotoxicity data was used. The data set comprises 677 chemicals reporting their ecotoxicity in algae (569), daphnia (622), and fish (617). In this study, a total of 505 compounds for algae, 547 for daphnia, and 505 for fish toxicity were selected, excluding the inorganic and replicate compounds (Table S1 of Supporting Information). Acute toxicity data of organic chemicals to OECD recommended test species was considered as external test data. Accordingly, data sets describing experimental ecotoxicities of chemicals in another algae species (S. obliguue) and marine bacteria (V. f ischeri) were considered here49 (Table S2 of Supporting Information). In these data sets, the toxicity values (EC50) for algae and marine bacteria were the 50% effective inhibition concentration at 48 h and 15-min exposure, respectively.

H=

12 N∑(N∑ + 1)

m

∑ i=1

R i2 − 3(N∑ + 1) Ni

(1)

where Ni is the sample size of each m group, Ri is the sum of the ranks of the ith group, and NΣ = Σim= 1Ni is the total number of samples considered in the experiment. The H-statistics approximates a χ2 with m − 1 degrees of freedom if the null hypothesis of equal populations is true. If the test statistics is greater than the χ2 percent point function, the null hypothesis is rejected.58 This test estimates the value of χ2 statistics for the experimental end-points and, therefore, evaluates the probability that samples were drawn from two statistically different groups. When the 743

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

value of χ2 exceeds the critical value for the selected significance level, the two groups are deemed significantly different.59 2.4. EL-QSAR Modeling. An ensemble contains a number of base learners.60 Ensemble learning is able to boost weak learners making accurate predictions. Bagging and boosting constitute popular ensemble learning methods and are considered here for constructing the classification and regression QSAR models (DTB and DTF) to analyze the ecotoxicity data. Stochastic gradient boosting and bagging algorithms are implemented here for constructing the QSAR models (DTB, DTF). A brief description of these methods is given below. 2.4.1. DTB-QSAR Model. In DTB, stochastic gradient boosting improves the accuracy of a predictive function.61 The DTB algorithm creates a tree ensemble, and it uses randomization during the tree

drawn randomly from training data and used to predict the entire data. These generated models are then aggregated. Bagging when combined with the base learner reduces variance enhancing the performance.63 The DTFs gaining strength from the bagging technique use the out-ofbag data rows for model validation. This provides an independent test set without requiring a separate data set or holding back rows from the tree construction. The method uses surrogate splitters for handling missing predictor values. The stochastic element in the DTF algorithm makes it highly resistant to overfitting. The number of trees in DTF and the depth of individual trees are the method’s parameter, which needs to be adjusted for optimal model selection. 2.5. Model Validation and Performance Criteria. The optimal architectures and parameters of the constructed classification and regression QSAR models were determined following both the internal and external validation procedures. Internal validation was performed using a V-fold cross-validation (CV) method, whereas for external validation, a separate validation (test) subset of the data was used, which was kept out during the training process.56 Optimal models were selected on the basis of the misclassification rate (classification) and mean squared error (regression) in the training and validation data.64 Predictive power of the constructed regression models for external sets was evaluated using the OECD recommended65 validation criteria (Q2F3) parameter.66 In Q2F3, the denominator is calculated on the training set65 as follows: n

2 Q F3 =1−

[∑i =ext1 (yî − yi )2 ]/next n

2 [∑i =Tr1 (yi − yTr ̅ ) ]/n Tr

(3a)

where yi and ŷi are the observed and model calculated value of the dependent variable, yT̅ r is the average of the experimental values of the training set, and nTr and next are the number of compounds in training and external sets. The performance of the binary classification QSAR models were assessed in terms of the sensitivity, specificity, and accuracy of prediction, computed as follows:52

sensitivity =

TP TP + FN

(3b)

specificity =

TN TN + FP

(3c)

accuracy =

creations (Figure 2a). In training set {x,y}, after each iteration, F represents the sum of all trees built so far: (2a)

where m is the number of trees in the model. Regardless of the lossfunction, the trees fitting the gradient on pseudoresiduals are regression trees trained to minimize mean squared error (MSE). The regularization parameter is the number of gradient boosting iterations and is achieved by shrinkage, which consists of modifying the update rule as follows: Fm(x) = Fm − 1(x) + υ·γmhm(x).

0100 mg/L) compounds based on EEC criteria16 and to predict the ecotoxicity of the chemicals. Architecture and parameters of the constructed QSAR models were optimized using the 10-fold CV procedure. For the 10-fold CV, data were split into 10 folds, and the single fold of data was predicted each time using a model built on the other remaining 9 folds, until all of the 10 folds were predicted. Thus, each time, a model is constructed and tested with an unseen data set. CV performs reliable and unbiased testing on data sets.52 The models were then applied to predict the toxicity of chemicals in other test species. The results obtained are discussed below. 3.1.1. Classification QSAR Modeling. Classification modeling was performed to categorize the chemicals as toxic and nontoxic. Accordingly, EL-based QSAR models (DTB and DTF) were developed. Internal and external validations were performed to determine the optimal architecture and the model parameters. A 10-fold CV was adopted for internal validation, whereas a separate subset of data (15%) was used for external validation. Average misclassification rate (MR) values of 10 runs in CV for DTB and DTF were 14.25% and 14.06%, respectively. The results suggest comparable prediction accuracies of both the QSAR models and show no obvious overfitting of data. The optimal DTB-QSAR model has a total number of trees in series, a maximum depth of the tree, the number of average group splits, and shrinkage factor values of 384, 6, 257.3, and 0.01, respectively. The two-category classification DTB model fits the logit values (probability). The shrinkage factor improves the predictive accuracy of a DTB series.73 In the DTF model, the total number of trees in series, maximum depth of a tree, and the number of average group splits were 190, 16, and 39.1, respectively.

where xi is a raw vector of molecular descriptors for a particular ith chemical. The fact that the value of hi is greater than the critical h* value indicates that the structure of the chemical substantially differs from those used for the calibration. Therefore, the chemical is located outside the optimum prediction space. The h* value can be calculated71 as follows:

h* =

3(p + 1) n

(4b)

where p and n represent the number of variables in the model and the number of training compounds, respectively. A leverage value greater than 3i/j is considered large and indicates that the predicted response is the result of an extrapolation.56 To visualize the AD based on both leverage and standardized residuals, the Williams plot is used. It is a plot of standardized residuals versus leverage values with two boundary lines, the first for outliers and the second for high leverage compounds.

3. RESULTS AND DISCUSSION Frequency distribution pattern of the experimental ecotoxicity values (−log EC50 or −log LC50 (mmol/L)) of chemicals pertaining to different test species used for the QSAR modeling are shown in the histograms (Figure 3a), which show a nearly

Figure 3. Histograms of the (a) toxicity values of the chemicals and (b) TSI values for the entire data sets in all of the considered test species.

normal distribution pattern of the toxicity values. The radar chart (Figure 1), a graphical method that displays multivariate data in the form of a two-dimensional chart shows that the compounds used in our data sets covered a sufficiently large chemical space.

Table 1. Correlation Coefficient and Number of Data Used for Both Species between the Toxicities of Two Speciesa

a

test species

P. subcapitata (algae)

S. obliguue (algae)

D. magna (daphnia)

O. latipes (fish)

V. f ischeri (bacteria)

P. subcapitata S. obliguue D. magna O. latipes V. f ischeri

1.000 (330) −0.227 (35) 0.014 (330) 0.051 (330) −0.164 (79)

1.000 (35) −0.398 (35) −0.183 (35) 0.199 (35)

1.000 (426) 0.012 (334) −0.194 (79)

1.000 (334) 0.026 (79)

1.000 (79)

The bold-faced value is significant (p < 0.05). 745

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

encodes the molecular distance edge between all secondary carbons. A molecular distance edge is defined as the through bond distance from atom A to atom B. This descriptor encodes molecular size and branching information of the molecule by taking into account sp3 hybridized C atoms.81 PPSA-1 and PNSA-1 are the charge partial surface areas.82 These descriptors provide information on atomic charges related to the whole molecule, weighted partial charge relative to surface areas, and fractional partial charges relative to surface areas81 and encode information on proton acceptor and donor sites. The nAtomP descriptor of compounds accounts for the stearic hindrance effect, which influences their transport properties through a biological system.83 The performance parameters of DTB and DTF models for the training, test, and complete data are presented in Table 2. The

The classification QSAR models were developed using nine descriptors belonging to topological such as a chi valence path descriptor of order 2 (VP-2), molecular distance edge between all secondary carbons (MDEC-22), nhigh lowest polarizability weighted BCUTS (BCUTp-1I), and directional WHIM, weighted by unit weights (Wlambda2.unity,WL-2U; Wlambda3.unity,WL-3U); electronic such as sum of the solventaccessible surface areas of all positively charged atoms (PPSA-1) and sum of the solvent-accessible surface areas of all negatively charged atoms (PNSA-1); and constitutional, such as the logarithmic form of octanol−water partition coefficient calculated by an atomic method (XLogP) and the number of atoms in the largest pi-system (nAtomP). Contribution of the selected descriptors in classification models (Figure 4a) ranged between

Table 2. Classification Result for Ecotoxicity Prediction of Chemicals in Toxic and Nontoxic Classes by EL-QSAR Modelsa model/subsets

class

total cases

Training Set (P. subcapitata) DTB 1 360 2 68 total 428 DTF 1 360 2 68 total 428 Test Set (P. subcapitata) DTB 1 65 2 12 total 77 DTF 1 65 2 12 total 77 Complete Set (P. subcapitata) DTB 1 425 2 80 total 505 DTF 1 425 2 80 total 505

Figure 4. Contribution of the selected descriptors in (a) classificationQSAR and (b) regression-QSAR models.

24.04%−100% (DTB) and 36.61%−100% (DTF). The discriminating descriptors in each model were determined in view of their importance in the corresponding model. The importance of the independent variables in each model was determined using the difference in MR calculated using actual data values of all predictors and those computed through random rearrangement values of each predictor. In all of the models, XLogP exhibited the highest contribution. XlogP calculated using an atom type prediction method describes the affinity of the compounds for partitioning the biological membranes.74 Chemicals with larger LogP values have higher ability to permeate the cell membrane and more easily interact with its target in the organism.75 The topological descriptors (VP-2, MDEC-22, BCUTp-1I, WL-2U, and WL-3U)) describe the molecules according to their size, branching, flexibility, and shape.76 BCUT are the weighted Burden matrix descriptors,77 which consider the connectivity and atomic properties of a molecule. In BCUTp-1I, the polarizability weighing scheme is employed.78 This descriptor will return the highest and lowest eigen values for the polarizability descriptor in a single array list and works with the H-depleted molecule. These descriptors describe the surface distribution of mass and polarizability of atoms.79 WL-2U and WL-3U are WHIM (weighted holistic invariant molecular) descriptors.80 These descriptors capture relevant molecular 3D information.80 The descriptor MDEC-22

a

sensitivity (%)

specificity (%)

accuracy (%)

98.09 100.00

100.00 98.09

98.36 98.36

99.72 100.00

100.00 99.72

99.77 99.77

95.52 90.00

90.00 95.52

94.81 94.81

95.52 90.00

90.00 95.52

94.81 94.81

97.70 98.59

98.59 97.70

97.82 97.82

99.07 98.70

98.70 99.07

99.01 99.01

1, toxic; 2, nontoxic.

MR values yielded by DTB and DTF models were 2.18% and 0.99% in complete data, respectively. The overall accuracy of the training and test sets were 98.36%, 94.81% for DTB and 99.77%, 94.81% for DTF model. Further, the results showed (Table 2) that the sensitivity and specificity values for DTB and DTF classification were more than 97% in complete data. For regulatory purposes, a classification model should be connected with high sensitivity.84 A higher specificity of the classification model represents its ability to recognize the false positive compounds.39 It is evident that both the DTB and DTF models yielded excellent results. The average gain values in DTB and DTF models ranged between 0.957 and 1.361 and 1.114−1.765, respectively, in validation data. The gain values show the ability of the model for picking out the best of the cases. The gain of 1 means nonselective targeting.63 The performance criteria parameters (sensitivity, specificity, and accuracy), and values of average gain for the classification QSAR models suggest that both the selected 746

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

Table 3. Performance Parameters for EL-QSAR Models for Ecotoxicity Prediction of Chemical Compounds experimental/model

subsets

mean

SDa

MAE

RMSE

R2

Q2F3

experimental

training (P. subcapitata) test (P. subcapitata) complete (P. subcapitata) D. magna O. latipes V. fischeri S. obliguue training (P. subcapitata) test (P. subcapitata) complete (P. subcapitata) D. magna O. latipes V. fischeri S. obliguue training (P. subcapitata) test (P. subcapitata) complete (P. subcapitata) D. magna O. latipes V. fischeri S. obliguue

1.55 1.57 1.55 1.63 1.45 1.16 0.69 1.53 1.53 1.53 1.54 1.55 1.26 1.28 1.56 1.57 1.57 1.59 1.60 1.31 1.25

1.22 1.16 1.21 1.13 0.90 0.71 0.67 1.01 0.85 0.98 0.94 0.81 0.63 0.69 0.92 0.73 0.89 0.84 0.74 0.54 0.61

0.24 0.43 0.28 0.52 0.47 0.34 0.60 0.30 0.50 0.34 0.52 0.47 0.33 0.57

0.34 0.56 0.39 0.67 0.61 0.43 0.71 0.40 0.64 0.46 0.67 0.59 0.45 0.67

0.946 0.793 0.918 0.654 0.575 0.661 0.672 0.940 0.753 0.905 0.656 0.605 0.648 0.689

0.791 0.897 0.696 0.753 0.877 0.657 0.726 0.861 0.698 0.770 0.865 0.695

DTB-QSAR

DTF-QSAR

a

SD: standard deviation.

(DTB); and 0.71 and 0.14 and 0.661 and 0.941 (DTF). These values are comparable to the results obtained when establishing the models in the training (DTB, 0.11 and 0.946; DTF, 0.16 and 0.940) and test (DTB, 0.31 and 0.793; DTF, 0.41 and 0.753) phases. A model is considered acceptable90 when the value of R2cv exceeds 0.5. The results indicate that both the models herein investigated are robust. Further, the results of the external validation of these models (Table 3) suggest that in all the cases, Q2F3 values were above its threshold (0.6).65 Consonni et al.66 demonstrated that results obtained by Q2F3 are independent of the prediction set distribution and sample size, hence independent of the samples chemical space. Moreover, according to the criteria proposed by Eriksson et al.91 the difference between R2 (training) and R2 (validation) should not exceed 0.3. As our models fulfill these criteria and also positively pass internal and external validations, these were applied to predict the toxicity of new, untested chemicals. The regression QSAR models were based on topological (chi simple path descriptor of order 1, SP-1; molecular distance edge between all secondary and tertiary carbons, MDEC-23; molecular distance edge between all tertiary carbons, MDEC33; autocorrelation descriptor weighted by scaled atomic mass, ATSm5; autocorrelation descriptor weighted by polarizability, ATSp5; and topological polar surface area, TopoPSA) and constitutional (XLogP; number of H-bond donors, nHBDon) descriptors. Contribution of the selected descriptors in constructed QSAR models is shown in Figure 4b. In DTB, among the eight selected descriptors, the contribution of XLogP was the highest (100%), whereas in DTF, SP-1 has the highest contribution. Compounds with a larger XLogP value have a higher tendency to partition into the cell and interact with the phospholipid membrane. A greater degree of H-bonding increases the polarity of the molecule and reduces toxicity.92 Distribution of a property along the topological structure of the molecule is explained by the 2D autocorrelation descriptors. TopoPSA of a molecule affects molecular transportation through membranes and therefore allows an estimation of the apparent

models are fully capable of discriminating the toxic and nontoxic chemicals. An in-depth investigation of the classification results (DTB) suggest that in complete data (P. subcapitata), 1 and 10 compounds, respectively, were misclassified as FP (imide) and FN (mainly comprising imides, aromatic amines, ester, hydrazine, neutral organics, phenols, and surfactant nonionic). Similar groups of compounds were also misclassified by the QSAR model applied to other external data sets (D. magna, O .letipes, V. f ischeri, and S. obliguue). The mode of action of imides is likely narcosis via membrane disruption.85 It can disrupt synthetic membranes. The aromatic compounds may act as redox cyclers and exert oxidative stress86 through generating highly reactive oxidants, such as the hydrogen peroxide and hydroxyl radical. Penetration of the substituted benzenes through the cell membrane and their electronic interactions with the active sites of action is mainly related to their toxicity in algae.87 In aromatic amines, bulky ortho substituted amines may force the amino moiety out of the aromatic plane, and this will reduce the stabilization of the developing charge when it starts to interact with the positively charged group. This charge may affect the interaction with membranes and influence the partitioning behavior.88 Hydrophobicity affects the penetration of a cell’s membrane by phenol and thus enhances the toxicity.89 The toxicity of surfactant is often related to the size of the hydrophobic component (number of carbons) or the number of repeating hydrophilic components (ethoxylates). 3.1.2. Regression QSAR Modeling. QSAR models were developed to predict the toxicity (−log EC50 mmol/L) of chemicals using the EL based modeling methods (DTB and DTF). Distribution of the selected descriptors for regression modeling is shown in the radar chart (Figure 1). A 10-fold CV was adopted to determine the optimal architecture and the model parameters, using the criterion of minimum MSE in the training and validation sets. The average values (10 runs) of MSEs and R2cv in internal validation and training data for the proposed QSAR models were 0.56 and 0.11 and 0.709 and 0.941 747

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

volume of distribution in the body.93 The constitutional descriptors reflect the molecular composition of a compound without any information about its geometry.94 The optimal QSAR (DTB and DTF) models have the total number of trees in series, maximum depth of the tree, and the number of average group splits 170 and 221; 7 and 15; and 376.6 and 150.8, respectively. The DTB model parameters were optimized using Huber M-regression loss function, which makes it highly resistant to the outliers.62 The DTF algorithm uses an out-of-bag validation procedure. The optimal DTB and DTF QSAR models were applied to the test and complete data sets. The constructed DTB and DTF models explained 89.49% and 85.78% variance in complete data. Proportion of variance explained by the model variables is the best single measure of how well the predicted values match the actual values. A model predicting exactly matching values with measured ones would explain 100% variance in data. The two models yielded MSE and R2 values of 0.15 and 0.918 (DTB) and 0.21 and 0.905 (DTF) in complete data. Values of the performance criteria parameters (R2, MAE, and RMSE) yielded by the QSAR models in training, test, and complete data are presented in Table 3. For a predictive QSAR model, the value of R2 in the external validation (test) set should be more than 0.5.90 From the results (Table 3), it is evident that the two models yielded considerably low RMSE and MAE values in training, test, and complete data. MAE measures the average magnitude of the error in a set of predictions, without considering their direction. Further, the experimental and model predicted values (training and test data) of the end-point are plotted in Figure 5. Both the measured and model (DTB and DTF) predicted toxicity values showed a close pattern of variation suggesting the adequacy of both the models. Plots of the residuals and model predicted toxicity values (Figure 6) of the chemicals exhibited a

random distribution95 pattern, which suggest that the models fitted the data well.

Figure 6. Plot of the residuals and model predicted values of the toxicity (−log EC50) values of chemicals in training and test set (a) DTB-QSAR and (b) DTF-QSAR models.

An investigation of the prediction results of the QSAR models revealed that most poorly predicted compounds, for which the difference between the experimental and predicted values of the end-point is more than 1.5 log unit, are 1 compound in model building data (thiols), 16 in daphnia (neutral organics, phenol, aliphatic amine, vinyl/allyl halide and alcohol, benzyl alcohol, ester, thiols, and hydrazine), 7 in fish (phenol, benzyl alcohol, neutral organic, aldehyde, acrylamide, and acrylates), and 1 in S. obliguue (aromatic amine). However, none of the chemicals violated this criterion in the case of bacteria data. Aromatic compounds have higher affinity for the attack of nucleophiles and reaction with reducing agents;86 hence, the benzene derivatives are representative of electrophilic toxicants, such as narcosis. Neutral organic chemicals are nonionizable and nonreactive and act via simple nonpolar narcosis. This general narcosis is often referred to as baseline toxicity.96 3.2. Application of QSARs to Multiple Test Species. The EL-based QSARs (classification and regression) models developed using the algae (P. subcapitata) ecotoxicity data of diverse chemicals were applied to other test species, such as algae (S. obliguue), daphnia, fish, and bacteria for predicting the toxicity of chemicals. Both the developed QSAR models yielded comparable excellent binary classification (toxic and nontoxic) accuracies of 92.50% (S. obliguue), 93.05% (daphnia), 94.26% (fish), and 94.12% (bacteria) for DTB-QSAR and 92.50% (S. obliguue), 92.14% (daphnia), 93.27% (fish), and 94.12% (bacteria) for DTF-QSAR. The regression QSARs also performed well with the external test data yielding high correlation (R2) between the measured and model predicted values of the response and low MSE values of 0.672 and 0.51 (S. obliguue), 0.654 and 0.45 (daphnia), 0.575 and 0.37 (fish), and 0.661 and 0.18 (bacteria) for DTB-QSAR; and 0.689 and 0.45 (S.

Figure 5. Plot of the measured and model predicted values of the toxicity (−log EC50) values of chemicals in training and test set (a) DTB-QSAR and (b) DTF-QSAR models. 748

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

Table 4. QSAR Models from Different Studies Reported in the Literature model

species

RSM LR LR LR LR MLR LR MLR LR LR LR LR LR MLR LR MLR MLR DTB-QSAR

C. vulgaris P. subcapitata P. subcapitata P. subcapitata P. subcapitata P. subcapitata P. subcapitata P. subcapitata V. f ischeri P. subcapitata P. subcapitata V. f ischeri P. subcapitata P. subcapitata P. subcapitata P. subcapitata D. magna P. subcapitata D. magna V. f ischeri S. obliguue O.latipes P. subcapitata D. magna V. f ischeri S. obliguue O. latipes

DTF-QSAR

duration

N

chemical class

R2

ref

48 h 24 h 48 h 48 h 96 h 48 h 48 h 15 min 3d 72 h 15 min 48 h 48 h 48 h 72 h 48 h 72 h 48 h 15 min 48 h 96 h 72 h 48 h 15 min 48 h 96 h

91 26 6 48 7 24 12 20 24 28 30 27 15 7 8 35 97 330 426 79 35 334 330 426 79 35 334

organic chemicals nonpolar narcotics nonpolar narcotics nonpolar narcotics chlorophenols phenols nitriles benzoic acids polar narcotics anilines phenols anilines propargylic alcohols(primary) propargylic alcohols(secondary) propargylic alcohols(tertiary) benzo-(triazoles) benzo-(triazoles) diverse chemicals diverse chemicals diverse chemicals diverse chemicals diverse chemicals diverse chemicals diverse chemicals diverse chemicals diverse chemicals diverse chemicals

0.870 0.941 0.970 0.870 0.941 0.810 0.850 0.964 0.810 0.551 0.850 0.471 0.760 0.850 0.970 0.821 0.769 0.918 0.654 0.661 0.672 0.575 0.905 0.656 0.648 0.689 0.605

Cronin et al.97 Hsieh et al.28 Escher et al.30 Tsai and Chen 29 Chen and Lin 25 Lee et al.23 Huang et al.32 Lee and Chen 24 Vighi et al.31 Aruoja et al.5 Aruoja et al.5 Aruoja et al.5 Chen et al.20 Chen et al.20 Chen et al.20 Gramatica et al.26 Cassani et al.27 this study this study this study this study this study this study this study this study this study this study

Table 5 descriptors

algae-train (P. subcapitata)

algae-test (P. subcapitata)

daphnia (D. magna)

bacteria (V. f ischeri)

algae (S. obliguue)

fish (O. latipes)

(a) Range of the Selected Descriptors in Classification QSAR Models VP-2 0.22−13.45 0.14−8.28 MDEC-22 0.00−40.49 0.00−27.41 BCUTp-1l 2.80−9.58 2.76−7.63 WL-2U 0.27- 29.92 0.54−10.39 WL-3U 0.99−217.41 0.88−46.45 PPSA-1 46.88−1624.99 34.61−806.68 PNSA-1 20.88−619.40 29.23−428.07 XLogP −3.17−16.79 −3.17−8.89 nAtomP 0.00−23.00 0.00−22.00 out of AD 3 (b) Range of the Selected Descriptors in Regression QSAR Models SP-1 1.41−15.57 2.41−10.27 MDEC-23 0.00−37.29 0.00−16.45 MDEC-33 0.00−20.18 0.00−6.08 ATSp5 0.00−10618.07 134.21−4848.44 ATSm5 0.00−168.85 1.81 −61.23 TopoPSA 0.00−155.68 0.00 −86.28 XLogP −3.17−10.94 0.09 −9.38 nHBDon 0.00−6.00 0.00 −2.00 out of AD

0.22−13.45 0.00−40.75 2.80−9.58 0.27−29.92 0.88−217.41 34.61−1624.99 20.88−619.40 −3.17−16.79 0.00−23.00 3

0.78−5.56 0.00−14.00 3.49−8.04 0.55−3.04 0.88−10.31 57.74−458.53 49.70−393.24 0.19−5.75 0.00−13.00 2

1.34−4.41 0.00−6.08 3.59−7.16 1.46−2.86 2.60−5.37 62.06−267.29 94.08−381.64 1.10−4.73 7.00−13.00

0.22−13.45 0.00−40.75 2.93−9.58 0.27−29.92 0.88−217.41 52.72−1624.99 20.88−619.40 −3.17−16.79 0.00−23.00 2

1.41−15.57 0.00−37.29 0.00−20.18 0.00 −10618.07 0.00−168.85 0.00−170.37 −3.17−10.94 0.00−6.00 1

1.41−6.45 0.00−10.13 0.00−9.13 0.00−1613.86 0.00−69.99 0.00−112.30 0.82−5.75 0.00−2.00

3.39−6.04 0.00−5.85 0.00−9.13 318.48 −1137.65 2.17−156.84 20.23−106.51 1.10−4.73 0.00−1.00

1.41−15.57 0.00−20.18 0.00−9.83 0.00 −10618.07 0.00−168.85 0.00−155.68 −3.17−9.38 0.00−4.00

3.3. Comparison with Other Studies. Several QSAR methods for toxicity prediction of chemicals in different test species are reported in the literature (Table 4). Since the data sets used in earlier research are generally small considering particular group based chemicals and differ between various models, a direct comparison of our results with these studies is inappropriate. Also, earlier QSARs are mostly based on

obliguue), 0.656 and 0.45 (daphnia), 0.605 and 0.34 (fish), and 0.648, 0.20 (bacteria) for DTF-QSAR models. Overall, the results suggest that both the EL-QSAR models are capable of predicting the acute toxicity of a broad range of molecules in different test species belonging to different trophic levels, hence providing effective tools for regulatory toxicology. 749

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

univariate or multiple linear regression (MLR) approaches using toxicity data determined over different time durations. Moreover, most of these studies considered complex descriptors, and in several of these, prediction accuracies were not satisfactory, thus limiting the applicability of these models for acute toxicity prediction purposes. The EL based QSAR modeling approaches in the present study considering the large data set of structurally diverse chemicals yielded the best prediction accuracy in complete data of different test species of different trophic levels recommended for toxicity evaluation of chemicals by the OECD. This study demonstrated the applicability of the EL-QSAR models in toxicity prediction of diverse chemicals in various test species and can be used as effective tools in regulatory decision making and risk assessment. Reliable QSARs models are required for toxicity prediction in multiple test species of various trophic levels to ensure a comprehensive risk assessment of chemicals. 3.4. Applicability Domain of QSAR Models. Validation of the classification and regression EL-QSAR models for screening new chemicals was performed through an analysis of the AD using the methods based on the range of the descriptors in a training set and those on leverage values for each of the compounds. According to the first approach, the ranges of descriptors calculated for the compounds in training, test, and external sets (classification and regression) are shown in Table 5. The results depict that all compounds in training, test, and external sets (classification) are inside the AD of the proposed models, except 3 compounds in P. subcapitata (carbon tetrachloride, formamide, and 1,3,5-triazine-2,4,6-trithiol), 3 in daphnia (carbon tetrachloride, 1,3,5-triazine-2,4,6-trithiol, and 1,1,2,2-tetraphenylethane-1,2-diol), and 2 each in bacteria (dichloro-methane and chloromethane) and fish (carbon tetrachloride and 1,1,2,2-tetraphenylethane-1,2-diol), whereas in regression models, all of the compounds in all of the data sets, except 1 compound in daphnia (4,4-oxybis(benzenesulfonylhydrazine)) were within the AD range. However, in this method the property under investigation is neglected.35 On the basis of the second method, the calculated leverage (h) values for all of the compounds considered in the entire test species were below the critical value (0.102), except for 2 compounds in daphnia (2,4,6-tribromophenol and indeno[1,2,3-cd]pyrene) and 1 compound each in S. obliguue (2,4,6-tribromoaniline) and fish (2,4,6-tribromophenol). To visualize the AD based on both the leverage and standardized residual, the Williams plot is used (Figure 7). As can be seen, five significant outliers are detected by analyzing the Williams plots pertaining to the DTB and DTF models. The results indicate that the predicted response is the outcome of interpolation of the model and therefore may be reliable. However, the h-value being less than h* does not necessarily guarantee that the compound falls within the AD because compounds may be located outside the AD due to large values of standardized residuals (values more than 3 times the SD unit).

Figure 7. Williams plot (a) DTB-QSAR and (b) DTF-QSAR models.

QSAR models were developed to categorize chemicals as toxic and nontoxic and to predict the acute toxicities of chemicals in different test species. Finally, multispecies QSAR models, useful for the reduction of animal tests have been developed and externally validated. Predictions generated by the proposed QSAR models for all the studied chemicals can be used by regulators to support the use of weight of evidence and nontesting based approaches. The proposed EL-QSAR models can be used as tools for safety evaluation of chemicals in regulatory toxicology.

■

ASSOCIATED CONTENT

S Supporting Information *

Toxicity data of chemicals in five different test species (P. subcapitata, D. magna, O. latipes, V. f ischeri, and S. obliguue). This material is available free of charge via the Internet at http://pubs. acs.org.

■

4. CONCLUSIONS This article proposes new global and robust QSAR models for the prediction of the aquatic toxicity of diverse chemicals in multiple test species of different trophic levels. These models have been developed and validated on the basis of OECD principles for QSAR acceptance and regulation, are characterized by high external predictivity and wide applicability domain, and have been applied to screen algae, daphnids, fish, and bacteria toxicity of diverse industrial chemicals, without requiring the knowledge of mechanisms for the specific toxicity end points.

AUTHOR INFORMATION

Corresponding Author

*Tel: 0091-522-2476091. Fax: 0091-522-2628227. E-mail: [email protected], [email protected]. Funding

A Senior Research Fellowship was provided by the Council of Scientific and Industrial Research (CSIR), New Delhi, India to S.G. Notes

The authors declare no competing financial interest. 750

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

■

Article

(6) Wei, D., Kisuno, A., Kameya, T., and Urano, K. (2006) A new method for evaluating biological safety of environmental water with algae, daphnia and fish toxicity ranks. Sci. Total Environ. 371, 383−390. (7) Dom, N., Knapen, D., Benoot, D., Nobels, I., and Blust, R. (2010) Aquatic multi-species acute toxicity of (chlorinated) anilines: experimental versus predicted data. Chemosphere 81, 177−186. (8) Henegar, A., Mombelli, E., Pandard, P., and Péry, A. R. R. (2011) What can be learnt from an ecotoxicity database in the framework of the REACh regulation? Sci. Total Environ. 409, 489−494. (9) Cetin, A. K., and Mert, N. (2006) Growth rate of Scenedesmus acutus (Meyen) in cultures exposed to trifluralin. Pol. J. Environ. Stud. 15, 631−633. (10) Organisation for Economic Cooperation and Development (1984) OECD Guidelines for Testing of Chemicals, Guideline 201, Algal Growth Inhibition Test, OECD Publishing: Paris, France. (11) Jeram, S., Riego Sintes, J. M., Halder, M., Baraibar Fentanes, J., Sokull-Klüttgen, B., and Hutchinson, T. H. (2005) A strategy to reduce the use of fish in acute ecotoxicity testing of new chemical substances notified in the European Union. Regul. Toxicol. Pharmacol. 42, 218−224. (12) Katsumata, M., Koike, T., Nishikawa, M., Kazumura, K., and Tsuchiya, H. (2006) Rapid ecotoxicological bioassay using delayed fluorescence in the green algae Pseudokirchneriella subcapitata. Water Res. 40, 3393−3400. (13) Castano, A., Cantarino, M. J., Castillo, P., and Tarazona, J. V. (1996) Correlation between the RTG-2 cytotoxicity test EC50 and in vivo LC50 rainbow trout bioassay. Chemosphere 32, 2141−2157. (14) Kaiser, K. L. E., and Palabrica, V. S. (1991) Photobacterium phosphoreum toxicity data index. Water Poll. Res. J. Canada 26, 361−431. (15) Kaiser, K. L. (1998) Correlations of Vibrio f ischeri bacteria test data with bioassay data for other organisms. Environ. Health Perspect. 106, 583−591. (16) Tan, N. X., Li, P., Rao, H. B., Li, Z. R., and Li, X. Y. (2010) Prediction of the acute toxicity of chemical compounds to the fathead minnow by machine learning approaches. Chemom. Intell. Lab. Syst. 100, 66−73. (17) Prakash, J., Nirmalakhandan, N., Sun, B., and Peace, J. (1996) Toxicity of binary mixtures of organic chemicals to microorganisms. Water Res. 30, 1459−1463. (18) Xu, S., and Nirmalakhadan, N. (1998) Use of QSAR models in predicting joint effects in multi-component mixtures of organic chemicals. Water Res. 32, 2391−2399. (19) Ren, S., and Frymier, P. D. (2002) Estimating the toxicities of organic chemicals to bioluminescent bacteria and activated sludge. Water Res. 36, 4406−4414. (20) Chen, C. Y., Kuo, K. L., and Fan, J. W. (2012) Toxicity of propargylic alcohols on green algae Pseudokirchneriella subcapitata. J. Environ. Monit. 14, 181−186. (21) Cronin, M. T. D., and Schultz, T. W. (1997) Validation of Vibrio f ischeri acute toxicity data: mechanisms of action−based QSARs for non-polar narcotics and polar narcotic phenols. Sci. Total Environ. 204, 75−88. (22) Zhao, Y. H., Cronin, M. T. D., and Dearden, J. C. (1998) Quantitative structure-activity relationships of chemicals acting by nonpolar narcosis−theoretical considerations. Quant. Struct. Act. Relat. 17, 131−139. (23) Lee, Y. G., Hwang, S. H., and Kim, S. D. (2006) Predicting the toxicity of substituted phenols to aquatic species and its changes in the stream and effluent waters. Arch. Environ. Contam. Toxicol. 50, 213−219. (24) Lee, P.-Y., and Chen, C.-Y. (2009) Toxicity and quantitative structure−activity relationships of benzoic acids to Pseudokirchneriella subcapitata. J. Hazard. Mater. 165, 156−161. (25) Chen, C. Y., and Lin, J. H. (2006) Toxicity of chlorophenols to Pseudokirchneriella subcapitata under air-tight test environment. Chemosphere 62, 503−509. (26) Gramatica, P., Cassani, S., Roy, P. P., Kovarich, S., Yap, C. W., and Papa, E. (2012) QSAR modeling is not “push a button and find a correlation”: a case Study of toxicity of (benzo-)triazoles on algae. Mol. Inf. 31, 817−835.

ACKNOWLEDGMENTS

We thank the Director, CSIR-Indian Institute of Toxicology Research, Lucknow (India) for his keen interest in this work and for providing all of the necessary facilities.

■

ABBREVIATIONS AD, applicability domain; ASTER, the assessment tools for the evaluation of risk; ATSm5, autocorrelation descriptor weighted by scaled atomic mass; ATSp5, autocorrelation descriptor weighted by polarizability; BCUTS, weighted burden matrix descriptors; CDK, Chemistry Development Kit; CNN, computational neural networks; CV, cross-validation; DTB, decision treeboost; DTF, decision tree forest; EC50, half maximal effective concentration; ECOSAR, Ecological Structure−Activity Relationships; EL, ensemble learning; FN, false negatives; FP, false positives; GLP, good laboratory practices; K−W, Kruskal− Wallis; LC50, median lethal concentration; LR, linear regression; MAE, mean absolute error; MCASE, Multi Computer Automated Structure Evaluation; MDEC-22, molecular distance edge between all secondary carbons; MDEC-23, molecular distance edge between all secondary and tertiary carbons; MDEC-33, molecular distance edge between all tertiary carbons; MLR, multiple linear regression; MR, misclassification rate; MSE, mean squared error; nAtomP, number of atoms in the largest pi-system; nHBDon, number of H-bond donors; OASIS, optimized approach based on the structural indices set; OECD, Organization for Economic Cooperation and Development; PNSA-1, sum of the solvent-accessible surface areas of all negatively charged atoms; PPSA-1, sum of the solvent-accessible surface areas of all positively charged atoms; QSARs, quantitative structure−activity relationships; R2, squared correlation coefficient; REACH, Registration, Evaluation, Authorization and Restriction of Chemical Substances Regulations; RMSE, root mean squared error; SD, standard deviation; SMILES, simplified molecular input line entry system; SP-1, chi simple path descriptor of order 1; TN, true negatives; TOPKAT, toxicity prediction program; TopoPSA, topological polar surface area; TP, true positives; TSI, Tanimoto similarity index; VP-2, chi valence path descriptor of order 2; WHIM, weighted holistic invariant molecular descriptors; WL-2U, Wlambda2.unity; WL3U, Wlambda3.unity; XLogP, logarithmic form of octanol− water partition coefficient calculated by an atomic method; χ2, chi-square

■

REFERENCES

(1) Casalegno, M., Sello, G., and Benfenati, E. (2006) Top-priority fragment QSAR approach in predicting pesticide aquatic toxicity. Chem. Res. Toxicol. 19, 1533−1539. (2) OECD (2004) Test No. 202: Daphnia sp. Acute Immobilisation Test, OECD Guidelines for the Testing of Chemicals, Section 2, OECD Publishing: Paris, France. DOI: 10.1787/9789264069947-en. (3) OECD (2011) Test No. 201: Freshwater Alga and Cyanobacteria, Growth Inhibition Test, OECD Guidelines for the Testing of Chemicals, Section 2, OECD Publishing: Paris, France. DOI: 10.1787/ 9789264069923-en. (4) OECD (1992) Test No. 203: Fish, Acute Toxicity Test, OECD Guidelines for the Testing of Chemicals, Section 2, OECD Publishing: Paris, France. DOI: 10.1787/9789264069961-en (5) Aruoja, V., Sihtmäe, M., Dubourguier, H.-C., and Kahru, A. (2011) Toxicity of 58 substituted anilines and phenols to algae Pseudokirchneriella subcapitata and bacteria Vibrio f ischeri: Comparison with published data and QSARs. Chemosphere 84, 1310−1320. 751

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

(27) Cassani, S., Kovarich, S., Papa, E., Roy, P. P., vander Wal, L., and Gramatica, P. (2013) Daphnia and fish toxicity of (benzo)triazoles: validated QSAR models, and interspecies quantitative activity-activity modelling. J. Hazard. Mater. 258−259, 50−60. (28) Hsieh, S. H., Hsu, C. H., Tsai, D. Y., and Chen, C. Y. (2006) Quantitative structure−activity relationships (QSAR) for toxicity of nonpolar narcotic chemicals to Pseudokirchneriella subcapitata. Environ. Toxicol. Chem. 25, 2920−2926. (29) Tsai, K. P., and Chen, C. Y. (2007) An algal toxicity database of organic toxicants derived by a closed-system technique. Environ. Toxicol. Chem. 26, 1931−1939. (30) Escher, B. I., Bramaz, N., Mueller, J. F., Quayle, P., Rutishauser, S., and Vermeirssen, E. L. M. (2008) Toxic equivalent concentrations (TEQs) for baseline toxicity and specific modes of action as a tool to improve interpretation of ecotoxicity testing of environmental samples. J. Environ. Monit. 10, 612−621. (31) Vighi, M., Migliorati, S., and Monti, G. S. (2009) Toxicity on the luminescent bacterium Vibrio f ischeri (Beijerinck) I: QSAR equation for narcotics and polar narcotics. Ecotoxicol. Environ. Safe 72, 154−161. (32) Huang, C.-P., Wang, Y.-J., and Chen, C.-Y. (2007) Toxicity and quantitative structure−activity relationships of nitriles based on Pseudokirchneriella subcapitata. Ecotoxicol. Environ. Saf. 67, 439−446. (33) Cronin, M. T. D., and Schultz, T. W. (1998) Structure activity relationships for three mechanisms of action of toxicity to Vibrio f ischeri. Ecotoxicol. Environ. Safe 39, 65−69. (34) Schultz, T. W., Cronin, M. T. D., Walker, J. D., and Aptula, A. O. (2003) Quantitative structure activity relationship (QSARs) in toxicology: a historical perspective. J. Mol. Struct. (Theochem.) 622, 1− 22. (35) Cash, G., and Nabholz, V., Eds. (2001) ECOWIN v0.99gECOSAR Classes for MS Windows, US EPA OPPT- Risk Assessment Division, Washington, DC. (36) Moore, D. R. G., Breton, R. L., and MacDonald, D. B. (2003) A comparison of model performance for six quantitative structure activity relationship packages that predict acute toxicity to fish. Environ. Toxicol. Chem. 22, 1799−1809. (37) Tunkel, J., Mayo, K., Austin, C., Hickerson, A., and Howard, P. (2005) Practical consideration on the use of predictive model for regulatory purposes. Environ. Sci. Technol. 39, 2188−2199. (38) Reuschenbach, P., Silvani, M., Dammann, M., Warnecke, D., and Knacker, T. (2008) ECOSAR model performance with a large test set of industrial chemicals. Chemosphere 71, 1986−1995. (39) Cheng, F., Shen, J., Yu, Y., Li, W., Liu, G., Lee, P. W., and Tang, Y. (2011) In silico prediction of Tetrahymena pyriformis toxicity for diverse industrial chemicals with substructure pattern recognition and machine learning methods. Chemosphere 82, 1636−1643. (40) Wang, X., Sun, C., Wang, Y., and Wang, L. (2002) Quantitative structure-activity relationships for the inhibition toxicity to root elongation of Cucumis sativus of selected phenols and interspecies correlation with Tetrahymena pyriformis. Chemosphere 46, 153−161. (41) Kar, S., and Roy, K. (2010) First report on interspecies quantitative correlation of ecotoxicity of pharmaceuticals. Chemosphere 81, 738−747. (42) Das, R. N., Sanderson, H., Mwambo, A. E., and Roy, K. (2013) Preliminary studies on model development for rodent toxicity and its interspecies correlation with aquatic toxicities of pharmaceuticals. Bull. Environ. Contam. Toxicol. 90, 375−381. (43) Snelder, T. H., Lamouroux, N., Leathwick, J. R., Pella, H., Sauquet, E., and Shanker, U. (2009) Predictive mapping of the natural flow regimes of France. J. Hydrol. 373, 57−67. (44) Yang, P., Yang, Y. H., Zhou, B. B., and Zomaya, A. Y. (2010) A review of ensemble methods in bioinformatics. Curr. Bioinform. 5, 296− 308. (45) Hancock, T., Put, R., Coomans, D., Vander Heyden, Y., and Everingham, Y. (2005) A performance comparison of modern statistical techniques for molecular descriptor selection and retention prediction in chromatographic QSRR studies. Chemom. Intell. Lab. Syst. 76, 185− 196.

(46) Dietterich, T. G. (2000) Ensemble methods in machine learning. Lect. Notes Comput. Sci. 1857, 1−15. (47) Mahjoobi, J., and Etemad-Shahidi, A. (2008) An alternative approach for the prediction of significant wave heights based on classification and regression trees. Appl. Ocean Res. 30, 172−177. (48) Japanese Ministry of Environment (2010) Japan Ecotoxicity Tests Data, March 2010. (49) Zhang, X. J., Qin, H. W., Su, L. M., Qin, W. C., Zou, M. Y., Sheng, L. X., Zhao, Y. H., and Abraham, M. H. (2010) Interspecies correlations of toxicity to eight aquatic organisms: Theoretical considerations. Sci. Total Environ. 408, 4549−4555. (50) Steinbeck, C., Hoppe, C., Kuhn, S., Floris, M., Guha, R., and Willighagen, E. (2006) Recent developments of the chemistry development kit (CDK) − An open-source Java library for chemoand bioinformatics. Curr. Pharm. Des. 12, 2110−2120. (51) Chemspider. www.chemspider.com. (52) Singh, K. P., Gupta, S., and Rai, P. (2013) Predicting acute aquatic toxicity of structurally diverse chemicals in fish using artificial intelligence approaches. Ecotoxicol. Environ. Safe 95, 221−233. (53) Hastie, T., Tribshirani, R., and Friedman, J. (2001) The Elements of Statistical Learning: Data Mining, Interference and Prediction, SpringerVerlag, New York. (54) Benigni, R., Netzeva, T. I., Benfenati, E., Bossa, C., Franke, R., Helma, C., Hulzebos, E., Marchant, C., Richard, A., Woo, Y. P., and Yang, C. (2007) The expanding role of predictive toxicology: An update on the (Q)SAR models for mutagens and carcinogens. J. Environ. Sci. Health, Part C 25, 53−97. (55) Zhao, C. Y., Zhang, H. X., Zhang, X. Y., Liu, M. C., Hu, Z. D., and Fan, B. T. (2006) Application of support vector machine (SVM) for prediction toxic activity of different data sets. Toxicology 217, 105−119. (56) Singh, K. P., Gupta, S., and Rai, P. (2013) Predicting carcinogenicity of diverse chemicals using probabilistic neural network modeling approaches. Toxicol. Appl. Pharmacol. 272, 465−475. (57) Zhao, L., Dong, Y. H., and Wang, H. (2010) Residues of veterinary antibiotics in manures from feedlot livestock in eight provinces of China. Sci. Total Environ. 408, 1069−1075. (58) Lowry, R. (2013) Concepts and Applications of Inferential Statistics, Online resource http://faculty.vassar.edu/lowry/webtext. html, 1999−2013, (accessed Sep, 2013). (59) Tcheslavski, G., and (Louis) Beex, A. A. (2010) Effects of smoking, schizotypy, and eyes open/closed conditions on the γ1 rhythm phase synchrony of the electroencephalogram. Biomed. Signal Proces. 5, 164−173. (60) Ishwaran, H., and Kogalur, U. B. (2010) Consistency of random survival forests. Stat. Probabil. Lett. 80, 1056−1064. (61) Friedman, J. H. (2002) Stochastic gradient boosting. Comput. Stat. Data An. 38, 367−378. (62) Huber, P. (1964) Robust estimation of a location parameter. Ann. Math. Stat. 53, 73−101. (63) Singh, K. P., Gupta, S., and Rai, P. (2013) Identifying pollution sources and predicting urban air quality using ensemble learning methods. Atmos. Environ. 80, 426−437. (64) Singh, K. P., Basant, N., and Gupta, S. (2011) Support vector machine in water quality management. Anal. Chim. Acta 703, 152−162. (65) Chirico, N., and Gramatica, P. (2011) Real external predictivity of QSAR models: how to evaluate it? Comparison of different validation criteria and proposal of using the concordance correlation coefficient. J. Chem. Inf. Model. 51, 2320−2335. (66) Consonni, V., Ballabio, D., and Todeschini, R. (2009) Comments on the definition of the Q2 parameter for QSAR validation. J. Chem. Inf. Model. 49, 1669−1678. (67) Singh, K. P., Basant, N., Malik, A., and Jain, G. (2010) Modeling the performance of “up-flow anaerobic sludge blanket” reactor based wastewater treatment plant using linear and nonlinear approachesA case study. Anal. Chim. Acta 658, 1−11. (68) Netzeva, T. I., Worth, A. P., Aldenberg, A., Benigni, R., Cronin, M. T. D., Gramatica, P., Jaworska, J. S., Kahn, S., Klpoman, G., Marchant, C. A., Myatt, G., Nikolova-Jeliazkova, N., Patliwicz, G. Y., Perkins, R., Roberts, D. W., Schultz, T. W., Stanton, D. P., van de Sandt, J. J. M., 752

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753

Chemical Research in Toxicology

Article

(89) Michałowicz, J., and Duda, W. (2007) Phenols − sources and toxicity. Pol. J. Environ. Stud. 16, 347−362. (90) Ojha, P. K., Mitra, I., Das, R. N., and Roy, K. (2011) Further exploring r2m metrices for validation of QSPR models. Chemom. Intell. Lab. Syst. 107, 194−205. (91) Eriksson, L., Jaworska, J., Worth, A. P., Cronin, M. T. D., McDowell, R. M., and Gramatica, P. (2003) Methods for reliability and uncertainty assessment and for applicability evaluations of classificationand regression-based QSARs. Environ. Health Perspect. 111, 1361−1375. (92) Parvez, S., Venkataraman, C., and Mukherji, S. (2008) Toxicity assessment of organic pollutants: Reliability of bioluminescence inhibition assay and univariate QSAR models using freshly prepared Vibrio fischeri. Toxicol. in Vitro 22, 1806−1813. (93) Fatemi, M. H., and Ghorbannezhad, Z. (2011) Estimation of the volume of distribution of some pharmacologically important compounds from their structural descriptors. J. Serb. Chem. Soc. 76, 1003− 1014. (94) Todeschini, R., and Consonni, V. (2000) Hand Book of Molecular Descriptors, Vol. 11, pp1−668, Wiley-VCH, New York. (95) Singh, K. P., Ojha, P., Malik, A., and Jain, G. (2009) Partial least squares and artificial neural networks modeling for predicting chlorophenol removal from aqueous solution. Chemom. Intell. Lab. Syst. 99, 150−160. (96) Franks, N., and Lieb, W. (1990) Mechanisms of general anesthesia. Environ. Health Perspect. 87, 199−205. (97) Cronin, M. T. D., Netzeva, T. I., Dearden, J. C., Edwards, R., and Worgan, A. D. P. (2004) Assessment and modeling of the toxicity of organic chemicals to Chlorella vulgaris: development of a novel database. Chem. Res. Toxicol. 17, 545−554.

Tong, W., Veith, G., and Yang, C. (2005) Current status of methods for defining the applicability domain of (quantitative) structure-activity relationship. Altern. Lab. Anim. 33, 155−173. (69) Nikolova-Jeliazkova, N., and Jaworska, J. (2005) An approach to determining applicability domains for QSAR group contribution models: an analysis of SRC KOWWIN. ATLA. Altern. Lab. Anim. 33, 461−470. (70) Kovarich, S., Papa, E., Gramatica, P., Kovarich, S., Papa, E., and Gramatica, P. (2011) QSAR classification models for the prediction of endocrine disrupting activity of brominated flame retardants. J. Hazard. Mater. 190, 106−112. (71) Puzyn, T., Rasulev, B., Gajewicz, A., Hu, X., Dasari, T. P., Michalkova, A., Hwang, H. M., Toropov, A., Leszczynska, D., and Leszczynska, J. (2011) Using nano-QSAR to predict the cytotoxicity of metal oxide nanoparticles. Nat. Nanotechnol. 6, 175−178. (72) Lessigiarska, I., Cronin, M. T., Worth, A. P., Dearden, J. C., and Netezeva, T. I. (2004) QSARs for toxicity to the bacterium Sinorhizobium meliloti. SAR QSAR Environ. Res. 15, 169−190. (73) Friedman, J. H. (2001) Greedy function approximation: A gradient boosting machine. Ann. Stat. 29, 1189−1232. (74) Chau, Y. T., and Yap, C. W. (2012) Quantitative nanostructure− activity relationship modelling of nanoparticles. RSC Adv. 2, 8489− 8496. (75) Jiang, D. X., Li, Y., Li, J., and Wang, G. X. (2011) Prediction of the aquatic toxicity of phenols to Tetrahymena Pyriformis from molecular descriptors. Int. J. Environ. Res. 5, 923−938. (76) Gupta, L., Patel, A., Karthikeyan, C., and Trivedi, P. (2010) QSAR studies on dihydro-alkoxy-benzyl-oxopyrimidines (DABOs) derivatives, a new series of potent, broad-spectrum non-nucleoside reverse transcriptase inhibitors. J. Curr. Pharm. Res. 01, 19−25. (77) Burden, F. R. (1997) Chemically intuitive molecular index. Quant. Struct. Act. Relat. 16, 309−314. (78) Kang, Y. K., and Jhon, M. S. (1982) Additivity of atomic static polarizabilities and dispersion coefficients. Theor. Chim. Acta 61, 41−48. (79) An, L. K., Li, Z. C., Wu, X. W., Zou, X. Y., and Gu, L. Q. (2013) The activity against methicillin-resistant Staphylococcus aureus and quantitative structureactivity relationship (QSAR) study of azanaphthindolizinedione derivatives. Afr. J. Pharm. Pharmacol. 7, 1020− 1025. (80) Todeschini, R., and Gramatica, P. (1998) 3D QSAR in Drug Design (Kubinyi, H., Folkers, G., and Martin, Y. C., Eds.) Vol. 2, pp 355−380, Kluwer/ESCOM, Dordrecht, The Netherlands. (81) Serra, J. R., Thompson, E. D., and Jurs, P. C. (2003) Development of binary classification of structural chromosome aberrations for a diverse set of organic compounds from molecular structure. Chem. Res. Toxicol. 16, 153−163. (82) Stanton, D. T., and Jurs, P. C. (1990) Development and use of charged partial surface area structural descriptors in computer-assisted quantitative structure-property relationship studies. Anal. Chem. 62, 2323−2329. (83) Luan, F., Zhang, R. S., Zhao, C. Y., Yao, X. J., Liu, M. C., Hu, Z. D., and Fan, B. T. (2005) Classification of the carcinogenicity of N-nitroso compounds based on support vector machines and linear discriminant analysis. Chem. Res. Toxicol. 18, 198−203. (84) Fjodorova, N., Vracko, M., Novic, M., Roncaglioni, A., and Benfenati, E. (2010) New public QSAR models for carcinogenicity. Chem. Cent. J. 4, 1−15. (85) Rand, G. M., Wells, P. G., and McCarty, L. S. (2003) In Fundamentals of Aquatic Toxicology (Rand, G. M., Ed.) pp 3−67, Taylor & Francis, London. (86) Schmitt, H., Altenburger, R., Jastorff, B., and Schüürmann, G. (2000) Quantitative structure activity analysis of the algae toxicity of nitroaromatic compounds. Chem. Res. Toxicol. 13, 441−450. (87) Lu, G. H., Yuan, X., and Zhao, Y. H. (2001) QSAR study on the toxicity of substituted benzenes to the algae (Scenedesmus obliquus). Chemosphere 44, 437−440. (88) Ramos, E. U., Vaal, M. A., and Hermens, J. L. M. (2002) Interspecies sensitivity in the aquatic toxicity of aromatic amines. Environ. Toxicol. Pharmacol. 11, 149−158. 753

dx.doi.org/10.1021/tx400371w | Chem. Res. Toxicol. 2014, 27, 741−753